Low-latency detections of gravitational waves (GWs) from compact stellar binary coalescences are crucial to enable prompt follow-up observations to astrophysical transients by conventional telescopes, as demonstrated by the first joint GW and electromagnetic observations on July 17, 2017. Searching over the GW parameter space with the requirement of low-latency presents a computational challenge. This will become more severe considering denser sampling of the source space due to improving GW detector sensitivities. In our previous work, a low-latency matched filtering search method was developed, called Summed Parallel Infinite Impulse Response (SPIIR) filtering, which is suitable for parallelization, and an over 50x speedup of this method was achieved using Fermi-generation CPUs. In this paper, a multi-rate scheme for filtering, which reduces the computation time by a factor of several, is presented. The recent features in NVIDIA CPUs, namely the read-only data cache, warp-shuffle, and cross-warp atomic techniques, were exploited to improve the performance of filtering over previous GPU acceleration by a factor of 1.5 similar to 11x on a Maxwell-generation GPU, whereas on a Pascal-generation GPU, the hardwn be gained by employing these new techniques. This leads to an over 100are upgrade can bring along 7 similar to 11x speedup, and a further 1.5 similar to 2.5x speedup cax speedup for the multi-rate scheme using the Maxwell GPU, and an over 260x speedup using the Pascal GPU. This GPU-accelerated multi-rate scheme was incorporated into a low-latency search pipeline - the SPIIR pipeline - and an overall near-limit CPU usage reduction is expected. This IIR technique in general and its GPU acceleration technique here hold the potential to find applications in other signal processing fields, such as in image and radio astronomy data processing. (C) 2018 Elsevier B.V. All rights reserved.
