Optimizing General Design Objectives in Processor-Array Design

In this paper we present an improved search procedure for the General Parameter Method (GPM) 1]. Our procedure maps uniform dependence algorithms to application-specic processor arrays (PAs). It can optimize general design objectives with certain non-monotonicity properties, i.e., those that do not increase monotonically with the parameters. An example of such an objective is the minimization of the total completion time, including load and drain times. In contrast, earlier design methods can only deal with monotonic objectives. We present results for the matrix-product problem using this search technique ; application of the technique on the transitive-closure problem is presented elsewhere 2]. We also show that the parameters in GPM can be expressed in terms of the schedule vector ~ and allocation matrix S in the popular dependence-based method (DM), thereby allowing GPM to be used in DM for nding optimal designs for uniform dependence algorithms.