Applying fuzzy logic to codesign partitioning

ntil recently, it was possible to distinguish between software-1 and hardware-oriented approaches to design. 2,3 In software-oriented design, the designer implements the system in software and moves modules that do not meet certain requirements to hardware. Hardware-oriented design starts with a fully hardware implementation; the designer moves blocks to software according to delay times. Both, however, are based on the designer's knowledge and relatively simple cost functions. Hardware-software codesign allows simultaneous design of a system's hardware and software components, thus exploiting the advantages of both, optimizing performance and cost, and allowing easier redesign when needed. Incentive for research in this field has increased greatly since the development of hardware devices that can implement complex algorithms with excellent performance levels and moderate cost. In addition, RISC processor technology allows software implementations of several functions that previously required hardware implementations to meet time requirements. Some of the more interesting studies on codesign seek to establish an integrated design methodology supported by appropriate tools. Codesign includes the allocation of hardware and software, performance evaluation, validation, and synthesis. 4 One of the key items is evaluation of the system's cost and partitioning. We are striving to provide a possible solution for this area of codesign—main-ly in the embedded system arena. Our aim is to drastically reduce design costs and at the same time facilitate implementation choices by offering faster exploration of the possible alternatives. To achieve this, we propose a partitioning tool that also allows designers to evaluate the performance of single modules without having to actually implement them. Our tool benefits from the simultaneous use of soft computing techniques: fuzzy logic, genetic algorithms and neural networks. The fuzzy approach to system development falls into the category known as soft computing , 5 which is based on the lower computational cost inherent in imprecision and uncertainty. In particular, we chose fuzzy logic 6 because it can specify observed magnitudes in fuzzy terms—that is, in an imprecise way. It does not, in fact, use the traditional concept of membership or nonmembership in a given set, but refers to the concept of continuous degree of membership. Thus, we can use fuzzy logic to define an object as very good, slow, inexpensive, and so on. Unfortunately, a fuzzy approach normally supplies knowledge bases that are easy for a human to interpret but have no learning capacity. To overcome this limit, we combined a genetic algorithm technique with …