Dynamic precision management for loop computations on reconfigurable architectures

Reconfigurable architectures promise significant performance benefits by customizing the configurations to suit the computations. Variable precision for computations is one important method of customization for which reconfigurable architectures are well suited. The precision of the operations can be modified dynamically at run-time to match the precision of the operands. Though the advantages of reconfigurable architectures for dynamic precision have been discussed before, we are not aware of any work which analyzes the qualitative and quantitative benefits which can be achieved This paper develops a formal methodology for dynamic precision management. We show how the precision requirements can be analyzed for typical computations in loops by computing the precision variation curve. We develop algorithms to generate optimal schedules of configurations using the precision variation curves. Using our approach, we demonstrate 25%-37% improvement in the total execution time of an example loop computation on the XC6200 device.