High Performance Synthetic Aperture Radar Image Formation On Commodity Multicore Architectures

Synthetic Aperture Radar (SAR) image processing platforms have to process increasingly large datasets under and hard real-time deadlines. Upgrading these platforms is expensive. An attractive solution to this problem is to couple high performance, general-purpose Commercial-Off-The-Shelf (COTS) architectures such as IBM’s Cell BE and Intel’s Core with software implementations of SAR algorithms. While this approach provides great flexibility, achieving the requisite performance is difficult and time-consuming. The reason is the highly parallel nature and general complexity of modern COTS microarchitectures. To achieve the best performance, developers have to interweave of various complex optimizations including multithreading, the use of SIMD vector extensions, and careful tuning to the memory hierarchy. In this paper, we demonstrate the computer generation of high performance code for SAR implementations on Intel’s multicore platforms based on the Spiral framework and system. The key is to express SAR and its building blocks in Spiral’s formal domain-specific language to enable automatic vectorization, parallelization, and memory hierarchy tuning through rewriting at a high abstraction level and automatic exploration of choices. We show that Spiral produces code for the latest Intel quadcore platforms that surpasses competing hand-tuned implementations on the Cell Blade, an architecture with twice as many cores and three times the memory bandwidth. Specifically, we show an average performance of 39 Gigaflops/sec for 16-Megapixel and 100-Megapixel SAR images with runtimes of 0.56 and 3.76 seconds respectively.