A framework for low energy data management in reconfigurable multi-context architectures

In this paper, we present an approach to the problem of low energy data scheduling for reconfigurable architectures targeting digital signal processing (DSP) and multimedia applications. The main goal is the reduction of the energy consumed by these applications through the integration of the proposed data management framework within a compilation tool specifically conceived for these architectures. Two levels of on-chip data storage are assumed to be available in the reconfigurable architecture. Then, the data manager tries to optimally exploit this storage hierarchy by saving data transfers among on-chip and external memories, so reducing the energy consumption. To do that, specific algorithms for finding the data shared among the different computation kernels of the application have been developed. Also, a data placement and replacement policy has been designed. We also show how an adequate data scheduling could decrease the number of operations required to implement the dynamic reconfiguration of the system.