A dynamically constrained genetic algorithm for hardware-software partitioning

In this article, we describe the application of an enhanced genetic algorithm to the problem of hardware-software codesign. Starting from a source code written in a high-level language our algorithm deter-mines, using a dynamically-weighted fitness function, the most interesting code parts of the program to be implemented in hardware, given a limited amount of resources, in order to achieve the greatest overall execution speedup. The novelty of our approach resides in the tremendous reduction of the search space obtained by specific optimizations passes that are conducted on each generation. Moreover, by considering different granularities during the evolution process, very fast and effective convergence (in the order of a few seconds) can thus be attained. The partitioning obtained can then be used to build the different functional units of a processor well suited for a large customization, thanks to its architecture that uses only one instruction, Move.