A macro-flow graph generation technique, supporting automatic synthesis оf parallel programs for heterogeneous multicore processors

A new technique for generation of coarse-grained dataflow graphs is proposed. The graphs mentioned represent parallel programs, processing large blocked numerical matrices on heterogeneous multicore processors with explicitly-managed memory hierarchies. The technique proposed allows removing of overheads related to dynamic trasks generation and data dependency tracking, which arise when existing approaches to software library construction are used.