Parallel Structures for Asynchronous Microprocessors


Philip B. Endecott

Abstract

This paper considers the implementation of pipelining and superscalar parallelism in asynchronous processors. The performance of simple pipelines and superscalar structures is improved by asynchronous implementation. The organisation of general purpose processors is more complex: they include forwarding paths joining non-adjacent pipeline stages. Unfortunately the additional synchronisation required by forwarding paths is detrimental to performance, yet not having forwarding is also detrimental. A solution called conditional forwarding is proposed but it is found that conventional instruction sets do not permit conditional forwarding. To allow this technique to be used an alternative programming model called explicit forwarding is introduced. In a processor with explicit forwarding the destination to which the result of an instruction must be sent is indicated explicitly by the instruction; in contrast in a conventional instruction set the routing of the result of an instruction is deduced from the register specifiers of adjacent instructions. The paper concludes by describing an experimental processor called SCALP (Superscalar Asynchronous Low-Power Processor) which uses explicit forwarding.
Full paper (34K gzipped).