This tutorial implements the Back-Projection (BP) algorithm for Synthetic Aperture Radar (SAR) on AI Engines. The goal is to illustrate a larger system design example with several different computational workloads. BP makes a good example since it contains a half-dozen workloads with different characteristics that must be stitched together into an efficient data flow between DDR and the AI Engine over the NoC using GMIO. The tutorial illustrates the following AI Engine design concepts:
Using system partitioning methodology to plan and de-risk a larger system design example.
Using multi-rate AI Engine scheduling to coordinate processing of different workloads at the system level.
Implementing vectorized function approximation kernels for
cos(),sin()andsqrt()using new Vitis DSP Library IPs.Managing dynamic range across several AI Engine kernels and compute workloads.
Optimizing the throughput performance of a single AI Engine compute engine.
Illustrating how to parallelize the system level application across multiple compute engines to increase throughput.