Dynamic Reallocation of Functional Units in Superscalar Processors

Abstract

In the context of general-purpose processing, an increasing number of diverse functional units are added to cover a wide spectrum of applications. However, it is still possible to design custom logic adapted to a particular application that will perform far better than a processor. In an attempt to give it some adaptability, adding some reconfigurability can help improve performance. We propose to extend the possibilities of complex multifunction units by dynamically reallocating existing complex functional units as multiple simpler units. The fact that more than one simple unit is involved in the ”reconfiguration” process implies that the decision is more global and needs to be taken for a longer period of time. We show that in typical superscalar architectures, there are no major impediments to implementing such a decision scheme, and that on a specific reallocation opportunity we can achieve speedups of up to 56% over a mainstream superscalar processor and practically no losses.