The goal of this research is to provide a methodology for the automatic construction of instruction sets for specialized application domains. Specialized, programmable computers are useful in situations where a general purpose computer may be too slow or expensive for the application. This research augments current high-level synthesis tools which derive a computer implementation when given an algorithm or instruction set as input. The proposed technique takes benchmarks and a description of the computer data path as input and produces as output the instruction set that best uses the data path to run the benchmarks. Topics of research include reducing the dependency of the results on the compiler used, determining an adequate number of benchmarks needed to produce an instruction set well-matched to the application domain, and investigating alternative methods of instruction set derivation. The proposed tool, when coupled with current high- level synthesis systems, can provide a complete design system for quickly mapping a specialized problem domain into a high performance processor. This allows the realization of computation accelerators for a broad class of disciplines, including real time event processing, digital signal processing, and computational chemistry, physics, and biology.
