A GPU-based algorithm for astronomical image subtraction photometry

Abstract

Astronomical image subtraction photometry is an essential method in the discovery of new celestial or astronomical phenomena. However, the original algorithm for astronomical image subtraction photometry costs long time when dealing with large scale images. So in order to match the trend of large data volume in the field of astronomy and the requirement of realtime processing, an algorithm based on CUDA GPU is proposed to accelerate astronomical image subtraction photometry. This paper analyzes the overall performance of the original algorithm and migrates the template image convolution part which is the most time consuming into GPU. The implementation and experiment of the proposed algorithm is based on HOTPANTS package, which is an open-source tool for astronomical image subtraction photometry. The experiment shows that the proposed algorithm is almost 9 times faster than the original one in the template image convolution part. The overall performance is improved by 50% to 60% in comparison with the original algorithm with acceptable accuracy.