Hybrid Genetic Algorithm for Allocation Mapping in Processor Array Design.

In this paper a hybrid genetic method for processor arrays design dedicated to realization of linear algebra algorithm for banded matrices is presented. The proposed method is a modification of previous genetic algorithm which is characterized by few important advantages relative to well-known linear projection methods. The main disadvantages of this previous method are: long program runtime and problems with obtaining acceptable allocation mapping results for huge information dependency graphs. Linear projection methods don't allow obtaining better allocation mapping solutions but are characterized by shorter program runtime. New hybrid algorithm combines these both linear and genetic methods and merges their advantages. Summarizing, this new proposed method is characterized by: a shorter program runtime, better allocation mapping results in comparison with both previous methods, possibly allocation mapping for large input linear algebra banded matrices and possibility of defining the designed processor array structure.