Tatami: Dynamic CGRA Reconfiguration for Multi-Core General Purpose Processing
TimeWednesday, December 8th6:00pm - 7:00pm PST
LocationLevel 2 - Lobby
DescriptionCGRAs enable energy-efficient, highly configurable solutions for embedded system workloads. Most existing techniques have been proposed in the context of single core systems and are limited to benefit multi-core configurations.
To take advantage of multi-core systems, one needs to support dynamically offloading multiple kernels to a CGRA accelerator. As all of the combinations cannot be determined and programmed upfront, by adaptively programming the CRGA in run time, we exploit the acceleration opportunities. In this work, we introduce a novel CGRA programming method that limits the complexity of selecting the optimal configuration, while maintaining close-to-optimal performance results.