Abstract
Time-domain radio astronomy utilizes a harmonic sum algorithm as part of the Fourier domain periodicity search, this type of search is used to discover single pulsars. The harmonic sum algorithm is also used as part of the Fourier domain acceleration search which aims to discover pulsars that are locked in orbit around another pulsar or compact object. However porting the harmonic sum to many-core architectures like GPUs is not a straightforward task. The main problem that must be overcome is the very unfavourable memory access pattern, which gets worse as the dimensionality of the harmonic sum increases. We present a set of algorithms for calculating the harmonic sum that are more suited to many-core architectures such as GPUs. We present an evaluation of the sensitivity of these different approaches, and their performance. This work forms part of the AstroAccelerate project which is a GPU accelerated software package for processing time-domain radio astronomy data.
Introduction
In the search for pulsars—rapidly rotating neutron stars—astronomers look for faint, periodic radio pulses. Because these signals are often buried in noise, scientists use Fourier transforms to move into the frequency domain. However, a pulsar’s signal is rarely a single spike; its power is typically spread across multiple "harmonics" (integer multiples of the base frequency).
To detect these faint signals, astronomers use a harmonic sum algorithm to gather scattered power back into a single bin, significantly boosting the Signal-to-Noise Ratio (SNR). While conceptually simple, this process is computationally demanding for modern hardware like GPUs. As higher harmonics are summed, memory access patterns become increasingly erratic and non-contiguous, leading to massive performance bottlenecks and load-balancing issues. This research introduces a suite of GPU-accelerated algorithms designed to overcome these memory constraints.
Methodology
The study implemented and compared four different approaches to harmonic summing within the AstroAccelerate framework:
- Standard HRMS: The baseline "optimal" algorithm that performs exhaustive sums to find the best SNR.
- Max HRMS: A performance-oriented version that selects the maximum value within a range of possible peak locations to add to the partial sum.
- Freq. bin HRMS: The simplest approach, summing only bins that are exact integer multiples of the fundamental frequency, ignoring potential signal drifts.
- Greedy HRMS: A recursive approach that adds the value of a bin or its neighbor—whichever is larger—to the partial sum, effectively "tracking" the signal peak.
To evaluate these methods, the researchers used synthetic datasets consisting of artificial pulsar signals (modeled via a modified von Mises distribution) embedded in white noise, simulating 20-second observations with a 64-microsecond sampling time.
Performance and Sensitivity Results
The primary challenge in GPU porting is not raw mathematical operations, but the movement of data from memory to GPU cores. The results demonstrate that algorithmic approximations can significantly mitigate these bottlenecks with minimal loss in scientific accuracy.
As shown in the performance data, the Greedy HRMS on GPU achieved a staggering 704x speedup over its CPU counterpart and was 6.4x faster than the standard GPU implementation. The Freq. bin HRMS was the fastest overall, reaching a 10.1x speedup over the standard GPU version, though this speed comes at a cost to sensitivity.
Sensitivity analysis reveals that Greedy HRMS maintains a detection capability nearly identical to the standard "optimal" sum (often represented by the Presto baseline in literature). In contrast, the Freq. bin HRMS suffered a ~20% loss in SNR, and Max HRMS was the least sensitive, losing approximately 40% of the signal.
Conclusion
The "Greedy" algorithm represents a significant breakthrough for real-time radio astronomy. It provides an excellent middle ground, effectively tracking signal peaks as they shift in higher harmonics without the massive computational overhead of checking every possible combination. While the Freq. bin and Max versions offer extreme speed, they risk missing the faintest pulsars. The Greedy HRMS ensures that as next-generation radio telescopes generate unprecedented volumes of data, processing pipelines can maintain high sensitivity without being overwhelmed by computational demands.
Summary updated