GPU-accelerated synthetic aperture radar backprojection in CUDA

Pleasingly parallel algorithms such as filtered back-projection have been documented to enjoy significant speedups when ported to run on a graphics processor instead of a standard CPU. Presented here is a two-dimensional SAR backprojection implementation for a single GPU using the NVIDIA CUDA framework. Given that input range projections may be too large to fit in graphics memory, our implementation allows the partitioning of the data along range as well as aperture. We show the efficacy of some CUDA optimizations, and apply it to two public datasets, one of which has varying start frequency.