Hardware-accelerated consistent computing structure for signal processing and deep learning

A variety of typical signal processing algorithms and deep learning algorithms were analyzed and modularized from the calculation requirements level. The computing modules, which were suitable for hardware acceleration parallelly in the two types of applications were extracted. A consistent computing model for signal processing and deep learning was proposed, and a hierarchical processing element and arrayed processing structure were proposed based on the consistent computing model in which the control part and computation part were separated. By the software definition of different application computing processes, the consistent hardware-accelerated computation of signal processing and deep learning could be realized flexibly. Based on Zynq computing platform, the consistency computing model and computing structure were verified from two aspects of reconstruction efficiency and computing performance. The validation results indicate that software-defined reconfigurable computing structures based on consistency computing models have high computational performance and reconstruction efficiency.