Improved acceleration of the GPU Fourier domain acceleration search algorithm

We present an improvement of our implementation of the Correlation Technique for the Fourier Domain Acceleration Search (FDAS) algorithm on Graphics Processor Units (GPUs) (Dimoudi and Armour 2015; Dimoudi et al. 2017). Our new improved convolution code which uses our custom GPU FFT code is between 2.5 and 3.9 times faster the than our cuFFT-based implementation (on an NVIDIA P100) and allows for a wider range of filter sizes then our previous version. By using this new version of our convolution code in FDAS we have achieved 44% performance increase over our previous best implementation. It is also approximately 8 times faster than the existing PRESTO GPU implementation of FDAS (Luo 2013). This work is part of the AstroAccelerate project (Armour et al. 2002), a many-core accelerated time-domain signal processing library for radio astronomy