Elastic Multi-Context CGRAs
Omar Ragheb, Tianyi Yu, Rami Beidas, Jason Helge Anderson · 2022 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) · 2022
A key aspect of Coarse-Grained Reconfigurable Arrays (CGRAs) is dynamic reconfigurability, where multiple configurations, or contexts, are loaded into the CGRA to time-multiplex its resources. This feature allows the CGRA to accommodate larger applications without a significant increase in its size. Context switching is typically centralized, using the system clock to synchronously cycle through configurations simultaneously across CGRA resources. This approach is unable to efficiently accommodate variable-latency operations. Elastic CGRAs were proposed to handle such operations via an architecture that operates according to a dataflow paradigm. However, elastic solutions are single context by nature, which limits their applicability to smaller application kernels. Time-multiplexed multi-context and elastic CGRAs are thus naturally incompatible with one another. In this paper, we aim to overcome this incompatibility and propose an architectural framework that is capable of generating elastic CGRAs with multi-context support. Elastic primitives that traverse contexts in a distributed fashion are introduced. We also extend conventional mapping solutions to handle the new architectures. Finally, we evaluate the area and performance overhead for elastic multi-context CGRAs over single context ones with equal processing capacity.