Accelerating SPWVD : Efficient GPU-Based Implementation for Time–Frequency Analysis of LPI Radar Signals

Sammanfattning

Low Probability of Intercept (LPI) radar systems are designed to detect targets while minimizingthe risk of being detected themselves, placing strict requirements on signal processing tech-niques used in electronic warfare. Time–Frequency Analysis (TFA) plays a key role in automaticmodulation recognition of such signals, where both high resolution and computational efficiencyare essential. The Smoothed Pseudo Wigner–Ville Distribution (SPWVD) offers a favorable trade-off between resolution and cross-term suppression, but its high computational cost makes real-time application challenging. This thesis investigates the computational characteristics of the discrete SPWVD and explores itsefficient implementation on Graphics Processing Units (GPUs). A sequential formulation is firstanalyzed to identify inherent parallelism and performance bottlenecks. Based on this analysis,multiple CUDA-based implementations are developed, ranging from a naive parallelization tooptimized designs that exploit data parallelism, shared memory, and efficient memory accesspatterns. Performance is evaluated using execution time and achieved memory throughput, with compar-isons to practical hardware limits obtained through the STREAM benchmark. The results showthat the naive implementation severely underutilizes the GPU, while the optimized approachessignificantly improve performance. By restructuring the algorithm and applying frequency-domainconvolution, the final implementation achieves throughput close to the practical memory band-width limit of the GPU and reduces execution time by a factor of five. The study demonstrates that SPWVD can be effectively accelerated on modern GPUs, makingit viable for high-throughput signal processing. However, limitations related to memory transfersand scalability remain, highlighting opportunities for further optimization and future work