Abstraction and Acceleration of Tensor Processing for Element-Level Digital Arrays

This paper describes an abstraction of tensor operations applicable to digital array radars (DARs) which enables adaptive on-array processing to meet changing requirements. An application programming interface (API) is presented that allows multiple algorithms to be realized with the same high-level description, with tunable parameters including array geometry, desired region of reconstruction (RoR), etc. This abstraction facilitates algorithm scaling and reconfiguration while also providing an efficient implementation on available hardware. Results of the abstraction are presented for two different algorithms solving a two-dimensional near-field imaging problem consisting of a single emitter in a scene. Both algorithms are tested on both simulated and measured data for comparison. Additionally, we explored automation tools for field-programmable gate array (FPGA) code development. We have shown that an abstracted FPGA implementation is capable of achieving 3-4 orders of magnitude speedup over traditional computing resources.