查看文章

The advent of machine learning has resulted in the rapid development of machine learning accelerators that are capable of computing tensor operations efficiently. Specifically, these accelerators compute matrix-matrix multiplication, a key routine in linear algebra libraries and machine learning. While using the accelerators would result in high performance radar signal processing, the algorithms used often require significant redesign in order to efficiently map them on to existing machine-learning hardware. In this paper, we show that higher levels of abstraction facilitate the efficient mapping of array algorithms onto commercial-off-the-shelf (COTS) machine learning hardware that results in higher performance in terms of execution time and/or throughput. Furthermore, similar levels of abstraction can be used to design efficient implementations of ML algorithms for radar processing, resulting in improved radar …