Generating structured program instances with a high degree of locality

Memory hierarchy-consciousness is an important requirement for the design of high-performance programs. We describe a tool that supports the programmer in restructuring performance-critical code sections. The tool works with small program instances, which are obtained by fixing program parameters such as loop bounds, and rewriting the program as an operation sequence. The tool automatically reorders the operations for better locality, and respects data dependencies. It outputs the optimized program instance in a structured form. The user finally recognizes the locality-relevant structure and generalizes it to the program. The paper focuses an recent advances in the development of our method. In particular we introduce a hierarchical clustering scheme that highlights operation subsequences with much data reuse. The scheme is applied to the generation of structured optimized program instances in which the locality-relevant structure is easy to recognize. Experimental results are included