Bin recycling strategy for an accuracy-aware implementation of two-point angular correlation function on GPU

Abstract

Cosmological studies, in particular those relating to the large scale distribution of galaxies, have to cope with an extraordinary increase in data volume with the current and upcoming sky surveys. These usually involve the estimation of N-point correlation functions of galaxy properties. Due to the fact that the correlation functions are based on histogram construction, they have a high computational cost, which worsens with the ever growing size of the datasets and the standard sample. At the same time, correlation functions exhibit a high sensitivity to the accuracy of the estimation. Therefore, their implementations require maintaining a high accuracy within a reasonable processing time. GPU computing can be adopted to overcome the latter problem, but the standard implementation of the histogram construction on GPU lacks the appropriate accuracy for calculating the cosmological correlation functions. In this work, the bin recycling strategy is implemented and evaluated for the estimation of the Two-Point Angular Correlation Function. At the same time the lack of the appropriate accuracy in the calculation of diverse implementations of histogram construction on GPU is demonstrated. The bin recycling strategy for the Two-Point Angular Correlation Function outperforms other implementations while enabling the processing of a large number of galaxies. As a consequence of this work, an accuracy-aware GPU implementation of the Two-Point Angular Correlation Function is stated and evaluated to assure the correctness of the results.