Fitting Galaxies on GPUs

TL;DR

This paper explores implementing galaxy light profile fitting on GPUs to significantly accelerate the process, enabling higher quality fits for large observational datasets.

Contribution

It analyzes the suitability of galaxy fitting algorithms for GPU acceleration and presents a preliminary GPU implementation with promising speed-up results.

Findings

Approximate 10x speed-up over CPU implementation

Algorithms are well-suited for GPU's high memory bandwidth

Potential for improved fit quality and robustness

Abstract

Structural parameters are normally extracted from observed galaxies by fitting analytic light profiles to the observations. Obtaining accurate fits to high-resolution images is a computationally expensive task, requiring many model evaluations and convolutions with the imaging point spread function. While these algorithms contain high degrees of parallelism, current implementations do not exploit this property. With evergrowing volumes of observational data, an inability to make use of advances in computing power can act as a constraint on scientific outcomes. This is the motivation behind our work, which aims to implement the model-fitting procedure on a graphics processing unit (GPU). We begin by analysing the algorithms involved in model evaluation with respect to their suitability for modern many-core computing architectures like GPUs, finding them to be well-placed to take advantage of the high memory bandwidth offered by this hardware. Following our analysis, we briefly describe a preliminary implementation of the model fitting procedure using freely-available GPU libraries. Early results suggest a speed-up of around 10x over a CPU implementation. We discuss the opportunities such a speed-up could provide, including the ability to use more computationally expensive but better-performing fitting routines to increase the quality and robustness of fits.