Research On GPU Based Parallel Computing Techniques For Information Fusion Filtering Process

With theincreasing demands of the accuracy and real-time for information fusion system, its computing complexityhasincreased remarkably. The information acquired by multi-sensor data fusion system has many aspects that need to be processed carefully, such as format is inconsistent, dimensions is not uniform, too much interference and other issues.Therefore, there is a very strong demand for arithmetic improvement in practicalapplications, especially for parallelcomputing. The commonused parallel method is of CPU multi thread mode, using multi-core CPU or computer cluster method, the process is complicated and the cost is expensive, not effective.CUDA(ComputeUnifiedDevice Architecture) provides developersa friendly developmentenvironment to fully use GPU's computing power. GPU parallel computing method based on the CUDA architecture has the characteristics of low-cost hardware,easy programming,etc.It also makes a new requirement of getting high CPU/GPU cooperative computation efficiency. On the one hand, the load balance of threads on GPU should be preserved.on the other hand, the utilization of CPU should be kept highenoughwhile GPU is working. The major research is focused on using the CPU/GPU parallel computing multi-mode to realize a parallel filtering process methodforinformation fusion system.The main works are as follows:(1)With the analysison the CPU time for the information fuse system, the filter module is selected to be parallelized. Anddata structureof filter module is re-designedto match its realization on GPU.(2)Amulti thread parallel pattern is proposed on CPU to overcome the CPU/GPU computing power waste which is common in the synchronous computing pattern, which includesSystem main thread, data acquisition thread, thread of GPU control and GPU computing threads.Multi thread designmakes the CPUto sufficiently use its computing power without waiting for GPU idly, and the efficiency of the program is improved.(3) The CPU/GPU parallel computing solutionis realized, the statistics time costed by parallel process isanalyzed. The experiment results show that parallel results are correct and the expected parallel efficiency can be acquired. Moreover, a variety of program optimizing methods are put forward.