GPU acceleration of the frequency domain acceleration search for binary pulsars

Graphics processing units (GPUs) have been used to accelerate computation in a broad range of fields; this work presents a GPU-accelerated search for pulsars. Pulsars are highly magnetised neutron stars with extremely stable rotational periods. These periods can be accurately measured, which makes them exceptionally powerful reference tools in the field of astrophysics. Pulsars have very weak emissions, making them difficult to find. Most pulsars are found in large-scale surveys, which generate a large amount of data, and require extensive data processing. This work describes a GPU-based solution, with implications for real-time processing of pulsar search data. Pulsar astronomy uses radio telescope observations with high spectral and temporal resolution, which produce very large data sets and require intensive Digital Signal Processing. Large-scale pulsar surveys using next-generation radio telescopes such as the Square Kilometre Array (SKA), will have to be performed in real time as the volumes of raw data produced will be too large to be stored for an extended period. These computational requirements are compounded when searching for binary pulsars as their orbital motion makes them difficult to detect using classic periodicity searches. However, these rare pulsars are of great interest to physicists, as they allow us to test general relativity. Acceleration searches are the most common technique for detecting signals from binary pulsars that may be missed by standard search techniques. One of these, the frequency domain acceleration search (FDAS), mitigates the effect of orbital acceleration by correlating a matched template with the spectrum of a signal. This method has been shown to be more efficient than the alternative time domain acceleration search (TDAS)s. Even so, it is extremely computationally intensive to perform on a large scale. The existing implementation, Accelsearch, is run on a central processing unit (CPU), which limits its performance. We address this problem by creating a GPU port of the FDAS. An analysis of the fundamental calculations on which the FDAS is based informs the design of a fully asynchronous pipeline that exploits multiple levels of parallelism. This entails developing a novel technique for calculating Fresnel integrals, which increases the speed and numerical accuracy of the calculations, in both single- and double-precision. Furthermore, we develop a new estimate which improves the numerical accuracy of filter coefficients for accelerations close to zero. The GPU-accelerated pipeline achieves speeds 30 to 70 times faster than the existing serial CPU implementation. Our results clearly show that GPU acceleration is effective at reducing the cost of processing the FDAS component, to the point at which the SKA1-mid survey data could be searched in real time using 340 to 675 desktop GPUs from the Pascal generation