Programming transparency and portable hardware interfacing: towards general-purpose reconfigurable computing

Abstract

Despite enabling significant performance improvements, reconfigurable computing systems have not gained widespread acceptance: most reconfigurable computing paradigms lack (1) a unified and transparent programming model, and (2) a standard interface for integration of hardware accelerators. Ideally, programmers should code algorithms and designers should write hardware accelerators independently of any detail of the underlying platform. We argue that achieving portability and uniform programming with only limited loss of performance is one of the main issues that hinder the widespread acceptance of reconfigurable computing. To make reconfigurable computing globally more attractive, we suggest a transparent, portable, and hardware agnostic programming paradigm. For achieving software code and hardware design portability, platform-specific tasks are delegated to a system-level virtualisation layer that supports a chosen programming model-much in the same way platform details are hidden from users in general-purpose computers. Although an additional abstraction inherently brings overheads, we show that the involvement of the virtualisation layer exposes potential optimisations that compensate the overheads and bring additional speedups. As a case-study, we present a real design and implementation of a number of building blocks of such system and discuss the challenges involved in materialising the others.