Parallel processing techniques for the processing of synthetic aperture radar data on GPUs

This paper presents a design for parallel processing of synthetic aperture radar (SAR) data using one or more Graphics Processing Units (GPUs). Our design supports real- time reconstruction of a two-dimensional image from a matrix of echo pulses and their corresponding response values. Key to our design is a dual partitioning scheme that divides the output image into tiles and divides the input matrix into sets of pulses. Pairs comprised of an image tile and a pulse set are distributed to thread blocks in a GPU, thus facilitating parallel computation. Memory access latency is masked by the GPU´s low-latency thread scheduling. Our performance analysis quantifies latency as a function of the input and output parameters. Experimental results were generated with an nVidia Tesla C2050 GPU having maximum throughput of 1030 Gflop/s. Our design achieves peak throughput of 293 Gflop/s, which scales well for output image sizes from 2,048 × 2,048 pixels to 4,096 × 4,096 pixels. Higher throughput can be obtained by distributing the pulse matrix across multiple GPUs and combining the results at a host device.