A Dynamic Partial Reconfigurable Overlay Framework for Python

Dynamic Partial Reconfigurable (DPR) systems enable the exchange of system modules during runtime, and thus, an application-specific optimization of the system. However, the development of these systems, their modules, and the integration is a complex task. In this paper, we present our framework for DPR overlays. With our approach we enable a description of the overlay, and thus the hardware, via software. Based on the software description, our framework assembles the overlay in the field. Therefore, pre-implemented processing modules are loaded into reconfigurable partitions via DPR. The concept of DPR overlays is similar to pre-compiled libraries in the software domain, which dynamically loaded into memory during runtime. The framework is based on PYNQ, and a Yocto-based tool-flow. For the evaluation of our framework, we show the implementation of an artificial neural network.