|Other Abstract||Digital backend is an indispensable part of the radio telescope system which mainly responsible for the digitization of analog signal, processing and high-speed transmission of data. With the improvement of GPU performance and development tool chains, the signal processing core of digital backend has turned to a GPU-centric computing system The GPU-based pulsar backend system has higher development efficiency, lower development cost, more flexible configuration and more powerful signal processing capability than the traditional FPGA-based backend system.
The core algorithm of the pulsar backend system includes multi-channel digital filtering, RFI mitigation, de-dispersion processing and pulsar folding. Due to the hugeamount of computation, it is difficult to realize the real-time processing of massive data generated by ultra-wideband receiver equipment. The high-performance GPU cluster system can provide powerful computing resources, and theoretically solve the problem that the pulsar backend algorithm cannot be processed in real time. The parallelization of pulsar backend algorithm and the acceleration research of GPU-related algorithms are the key factor to the development of GPU-based digital backend, it is also crucial in the construction of pulsar data processing systems based on GPU clusters.
According to the actual problems encountered in the observation process of Xinjiang Astronomical Observatory and the needs of future observation and development, the thesis carried out the following research work.
(1) The research has realized high-speed polyphase channelizer technology basedon the new generation GPU and CUDA paralel architecture, and solved the problem that the large computational complexity and poor real-time signal processing capability of PFB algorithm. By designing the analysis polyphase FIR filter, we effectively reduced the spectrum leakage of Fourier transform, and investigated the method of suppressing spectrum leakages based on the window function, FIR filterand FFT. In this paper, the GPU parallel acceleration algorithm is proposed, which has a significant acceleration advantages compared with the CPU serial algorithm. The use of aligned and coalesced memory accesses maximizes the performance of the GPU algorithm. We applied the polyphase filer banks based on GPU, which fully utilized the multi-thread and multi-core paralel execution capability of the GPU, and greatly improved the real-time data processing performance of algorithm.(2)A new progress has been achieved in the RFI mitigation algorithms and techniques. We used the reference antenna methods to effectively suppress strong interference, preserve the integrity of the telescope data, and improved the sensitivity of the observation system An algorithm model of adaptive interference mitigation was proposed, and which automatically adjusts the filtering parameters according to the reference antenna signal, effectively eliminating the radio frequency interference mixed into the astronomical signal. We depth analyzed the RFI features, designed and optimized the adaptive RFI filters, and greatly improved the signal-to-noise ratio. In the design of GPU parallel algorithm, the kemel function use the GPU's global memory, shared memory, SP register and warp fast addition operation to optimized the parallel resources of the GPUs, and obtained an ideal speedup ratio, realized high-speed massive data processing. We studied the synchronization technique of the adaptive filter input signal. The RFI mitigation algorithm was test by simulation data and real data of Parkes ultra-wideband receiver.
(3) The GPU-based pulsar de-dispersion and folding parallel algorithm are implemented and studied, the high-performance parallel computing method is use to study the multi-thread processing of incoherent de-dispersion, coherent de-dispersion and pulsar folding algorithm. The parallelization acceleration scheme of the algorithm is proposed. By the coherent de-dispersion algorithm, we completely eliminated the dispersion effect and obtained a pulsar signal that close to real pulsar profile. The GPU algorithm used the cuFFT library to implement FFT and IFFT at high speed, optimized the GPU kemel of pulsar signal processing algorithm, and finally obtained satisfactory acceleration results. The of pulsar signal folding algorithm is implement, and the intensive calculation part of the algorithm is analyzed, the multi-thread task assignment, management and communication are studied, and the GPU's global memory, shared memory, constant memory and registers are utilized efficiently, thereby improved GPU resource utilization. In turn, the calculation time is been reduced and computational performance of the related algorithm is significantly improved.
In this thesis, we use NVIDIA's new generation GPU and general parallelcomputing architecture to study the algorithm of pulsar backend system systematically, introduced the hardware architecture and CUDA programming model of GPUs, and studied the parallel optimization methods of GPU algorithm. The polyphase fiter bank, adaptive RFI filtering, median filtering, incoherent de-dispersion, coherent de-dispersion and folding parallel algorithm are designed to solve the problem that the algorithm cannot calculated in real time on the CPU platform due to the huge amount of computation. The experimental results show that the designed parallel algorithm obtained several tens to hundreds of times acceleration ratio, and greatly improved the computational performance of the algorithm, fully demonstrated the huge advantages of the GPU computing platform. The related technologies have laid a foundation for the independent development of the pulsar digital backend system of the Xinjiang Astronomical Observatory.|