Implementing High-performance Intensity Model with Blur Effect on GPUs for Large-scale Star Image Simulation

Intensity model with blur effects are widely employed to accurately simulate the imaging process of a star simulator used for attitude determination and guiding feedback. The model is computationally intensive and the time requirements are proportional to the number of stars in the simulation, imposing great demands of computing power for realistic uses. This paper presents two star simulators using Graphic Processing Units (GPUs). We analyze the parallelism inherent in the intensity model and leverage a massive number of computing cores on GPU to efficiently exploit the fine-grain data parallelism. We first give a parallel simulator and discuss the performance trade-offs related to small amount of shared memory and the atomic operations on GPU. We then give the second simulator by adapting the first based on the characteristics of spatial locality with on-chip memory redesign. We analyze the balance between the kernel time and the non-kernel overhead in the two simulators and observe the inflection points in terms of two crucial model parameters. A selection table is given to choose between the two simulators. Benchmarks corresponding to the data parallelism are developed to fully evaluate the performance. The parallel simulator reports one to two orders of magnitude speedups with a maximum of 270 × compared to the widely-used sequential simulators and the average speedup is around 97 times. The adaptive simulator achieved up to 1.8×compared with the parallel one over the inflection point. The developed code is currently used for simulating complex star images in a realistic large-scale star simulator.