GPU-based Fourier Domain acceleration searching
This has been optimized for NVIDIA Kepler architecture. Might need additional tweaking for newer architectures.
The Fourier-Domain acceleration search technique is an optimized version of the algorithm described in http://adsabs.harvard.edu/abs/2002AJ....124.1788R. The algorithm described in this paper is also used in PRESTO CPU-accelerated implementation.
You might need additional tools such as PRESTO to perform de-dispersion and noise excision.
The algorithm involves the following procedures:
- Generating kernel template responses.
- Applying real-to-complex DFT of time-series of a specific DM trial.
- Whitened output of the DFT to remove red-noise.
- Dividing whitened signal in segments of equal length. Each segment has an overlapping region from the previous segment (using Overlap-and-Save method). (a) Generating `local Fourier amplitudes'. (b) Applying complex-to-complex DFT of segments. (c) Fourier Domain convolution of the segment with all kernel templates, producing a f - f dot plane. (d) Inverse complex-to-complex DFT of each convolved kernel template. (e) Calculating Fourier power of each harmonic. (f) Adding harmonics to fundamental harmonic (Harmonic summing). (g) Searching all folded harmonics powers (NOT IMPLEMENTED - See below for further detail)
NOTE that that acceleration searching requires candidate selection and a procedure to detect pulsars, including binary pulsars, from the folded Fourier power. In this implementation, though the search procedure was considered, the candidate selection was limited to the gener- ation of an empty list of potential candidates. The simplest method for selecting these candidates is to store all candidates above a certain S/N ratio and sort them accordingly. Several GPU-based sorting algorithms are available but these depends on the desired searching procedure.