Evaluating Multigauge Architectures for Computer Vision

Abstract

Heterogeneous computing is a technique for achieving high performance by providing a variety of different architectures to meet the needs of systems that are composed of tasks with widely different characteristics. Essential to the construction of heterogeneous systems is an understanding of the match between architecture and software and how that match can be used in deciding how to utilize the available computing resources. We present a theoretical framework, the PCI model, which defines corresponding characteristics of parallel programs and parallel architectures and defines the performance relationship between them in terms of these characteristics. We have encapsulated the concepts of the PCI model into RAW, a simulation environment that facilitates experimentation with the program/architecture relationship in terms of the PCI model. Using RAW, we have applied the PCI model to study the use of processor reconfigurable architectures (a type of heterogeneous system) in the context of computer vision applications. We present experimental results that demonstrate that these architectures perform better than static homogeneous architectures for such applications.