Fused Functional Units for Area-Efficient CGRAs

To solve the inefficiencies in general-purpose-computing architectures, many domain specific architectures have been proposed in the past decade. One of the promising architectures is coarse-grained reconfigurable arrays (CGRAs), which comprise functional units and a flexible interconnection network. In this paper, we present a framework that can generate area-efficient CGRAs via fused functional unit (FU) design. Our FUs support multi-precision for both fixed- and floating-point numbers. Our evaluation shows that the proposed fused functional units result in a 58%~60% decrease in FU area and an 11%~15% decrease across the whole CGRA when integrated with the state-of-the-art VectorCGRA framework [1], [2].