Automatic hardware accelerators reconfiguration through linearUCB algorithms on a RISC-V processor

Reconfigurable processors are hardware architectures that allow for the dynamic configuration of processing resources to optimize performance and power consumption, using partial reconfiguration to modify a portion of the design or update it without affecting the entire system. In this work, we present an automatic reconfiguration technique that leverages machine learning (ML) algorithms to automatically select the optimal configuration of a general-purpose hardware accelerator according to the workload and reconFigure the architecture at run-time. The problem is formulated as a Contextual Bandit (CB) case using the Linear Upper Confidence Bound (LinearUCB) algorithms and verified using the RISC-V Klessydra family cores as a case of study.