A Distributed Register File Architecture Based on Dynamic Scheduling for VLIW Machine
Yang Guo, Donglin Wang, Zijun Liu, Hongyu Meng · 2018
Distributed register files have the advantages of area and power compared with centralized register files in very long instruction word (VLIW) machine. But it may complicate the internetwork. A distributed register file architecture based on dynamic scheduling is proposed in our work. A portion of latency on internetwork is shifted to the latency among registers by latency transfer. And the decrease in the fanout of writing ports leads to a reduction in the power of the internetwork. To ensure the external compatibility of the distributed register file interface, the register was redirection designed. A dynamic scheduling strategy is designed to avoid conflicts between physical entities and logical entities that redirect registers. The optimization was evaluated on our in-house processor. The register file based on distributed dynamic scheduling reduces 15.48% of latency. The system frequency is raised from 1.19 GHz to 1.41 GHz. The power decreased by 27.42%.