Radio astronomical data pipeline consists of calibrations,RFI mitigation,coordinates translation,gridding,signal extraction,and et al.The most data intensive computing phase during the entire pipeline is gridding,which converts original data from irregular sampling space to regular grid space.As many new large radio telescopes have been established or are under construction,more and more observed data are being produced,which has become a great challenge for gridding capability.A number of methods are developed to provide efficient performances by utilizing heterogeneous architecture.However,existing proposals are designed for radio interferometers,or have not well considered large-scale data situation,or are customized to specify telescopes.To satisfying the data processing requirement of new-generation devices,there require fast algorithms to accelerate the process of gridding urgently.Here,combining with the features of large-scale telescopes and considering the advantages of CPU and GPU,we propose a CPU-GPU Hybrid Convolution-based Gridding(HyGrid)Algorithm.HyGrid speeds up convolution-based gridding by decreasing the search space of convolution and the calculation,as well as improving data access efficiency of device memory.Our contributions are focused on two part:(1)we utilize multicore CPU to rearrange sampled data,present a two level lookup table construction method to fast locate the required data of convolution and decrease calculation;(2)several optimization strategies are proposed to reduce unnecessary memory access and maximize the utilization of GPU,including configuring thread organization,using register and texture memory,and thread coarsening.Testing results demonstrate that the proposal is especially suitable for gridding large-scale data and can improve performance by up to 104.67 times compared to the traditional multi-thread CPU-based approach. |