Optimal datapath allocation for multiple- wordlength systems

Introduction: There has been significant recent research into multiplewordlength or multiple-precision systems, where datapaths are constructed from operators with different bit-width [1 – 3]. However, little research has been conducted [3, 4] into high-level synthesis for these systems. The use of multiple wordlengths has a significant impact on the traditional problems of high-level synthesis: scheduling, resource binding, and module selection [5]. This is the result of two factors. First, each computational unit of a specific type, for example ‘multiplier’, cannot be assumed to have an equal cost in a multiple-precision system [4]. Secondly, the choice of wordlength for an operation can impact on the latency of that operation. The existence of multiple wordlengths therefore complicates the resource binding problem, and also increases the interaction between operation binding and scheduling. To our knowledge, this Letter is the first formulation of the combined scheduling. resource binding and wordlength selection problem. The formulation is given in terms of an integer linear program (ILP), the solution of which yields optimal results with respect to the cost function. It is demonstrated that the area-based cost function can be used to provide bounds on the number of resource instances of each size and type, and an example is given. The notation f ( X ) for the range of a function f : X → Y is used in this Letter. | X | represents the cardinality of set X , and ∧ denotes logical AND.