high performance computing on graphics processing units: hgpu.org

We introduce and discuss a parallel SAR backprojection algorithm using a Non-Uniform FFT (NUFFT) routine implemented on a GPU in CUDA language. The details of a convenient GPU implementation of the NUFFT-based SAR backprojection algorithm, amenable to further generalizations to a multi-GPU architecture, are also given. The performance of the approach is analyzed in terms of accuracy and computational speed by comparisons to a "standard", parallel version of the backprojection algorithm exploiting FFT+interpolation instead of the NUFFT. Different interpolators have been considered for the latter processing scheme. The NUFFT-based backprojection has proven significantly more accurate than all the compared approach, with a computing time of the same order. An analysis of the computational burden of all the different steps involved in both the considered approaches (i.e., standard and NUFFT backprojections) has been also reported. Experimental results against the Air Force Research Laboratory (AFRL) airborne data delivered under the "challenge problem for SAR-based Ground Moving Target Identification (GMTI) in urban environments" and collected over circular flight paths are also shown.