Searches for weak gravitational-wave signals in the noisy data of ground-based interferometric detectors are notoriously difficult and computationally expensive. The following work explores the feasibility of improvements related to the details of software implementation using as an example a particular all-sky search for long-lived (continuous) gravitational waves emitted by rotating non-axisymmetric neutron stars (time-domain $\mathcal{F}$-statistic all-sky search (Astone et al., 2010)). In particular, we re-implement the majority of the code in the OpenCL framework, allowing it to execute on multi-device hybrid platforms. We also study the hypothesis that this code is not computationally bound, but memory bandwidth limited — by reducing the floating-point precision employed on various stages of the pipeline we note substantial performance gains. We report which stages may be computed with reduced accuracy without significant loss of precision to the results. The current implementation will allow its deployment on a much broader range of computing facilities.

众所周知，在地面干涉式探测器的噪声数据中搜索微弱的引力波信号非常困难，而且计算量很大。以下工作以旋转全非轴对称中子星（时域）发出的长寿命（连续）引力波的特定全天搜索为例，探讨了与软件实现细节相关的改进的可行性。$\mathcal{F}$统计全天搜索（Astone et al。，2010）。特别是，我们在OpenCL框架中重新实现了大多数代码，从而使其可以在多设备混合平台上执行。我们还研究了以下假设：该代码不受计算限制，但受内存带宽限制-通过降低流水线各个阶段所采用的浮点精度，我们注意到可观的性能提升。我们报告了哪些阶段可能会以降低的精度进行计算，而结果的精度却没有显着降低。当前的实现将允许将其部署在更广泛的计算设施上。