Algoritmos para alocação de recursos em arquiteturas reconfiguraveis

Recent work in reconfigurable architectures shows that they offer a better performance than general purpose processors (GPPs), while offering more flexibility than ASICs (Application Specific Integrated Circuits). A reconfigurable architecture can be adapted to implement different applications, thus allowing the specialization of the hardware according to the computational demands. In this work we describe an embedded systems project based on a reconfigurable architecture. We adopt an instruction set extension technique, where specialized instructions for an application are included into the instruction set of a GPP. These instructions correspond to sections of the application, and are executed in a dynamically reconfigurable datapath, added to the GPP’s hardware. The central focus of this theses is the resource sharing problem in the design of reconfigurable datapaths. Since the application sections are modeled as control/data-flow graphs (CDFGs), the CDFG merging problem consists in designing a reconfigurable datapath with minimum area. We prove that this problem is NP-complete. Our main contributions are two heuristic algorithms to the CDFG merging problem. The first has the goal of minimizing the reconfigurable datapath interconnection area, while the second minimizes its total area. Experimental evaluation showed that our first heuristic produced an average 26.2% area reduction, with respect to the most used method. The maximum error of our solutions was on average 4.1%, and some optimal solutions were found. Our second algorithm approached, in execution times, the fastest previous solution, and produced datapaths with an average area reduction of 20%. When compared to the best known area solution, our approach produced slightly better areas, while achieving an average speedup of 2500. The proposed algorithm also produced smaller areas, when compared to an industry synthesis tool.