Data path construction and refinement

Abstract

A system is described for the data path allocation problem in digital signal processor synthesis. The system, STAR, consists of three phases-preprocessing, data path construction (DPC), and data path refinement (DPR). The actions taken in each phase are described. The authors' contributions include the following: (1) theorems for the lower bound of the number of interconnections; (2) in the DPR phase, a more global view of the allocation problem is taken by ripping up and reallocating different types of objects simultaneously; (3) a novel technique to evaluate the binding quality of an object on the basis of a sharing of hardware resources which the object uses; (4) a method to judge the potential for upgrading a data path; and (5) an iterative improvement technique based on the idea of a relation network. The system currently supports the synthesis of architecture in linear topology and random topology. Parameters can be specified to explore different design alternatives and design space. Experiments on benchmarks show promising results.<>