Research And Application Of Tensor Decomposition And Reconstruction Method Based On GPU

Extracting the information contained in large-scale and multi-dimensional information data is a hot issue in the era of information and data technology.Among the existing method,tensor decomposition and reconstruction is an effective method for data analysis and information extraction for large-scale volume data.In the process of tensor decomposition,based on the principle of principal component analysis(PCA),the main features of information data are effectively retained and enhanced;in the process of tensor reconstruction,data can be quickly reconstructed at different scales,and can support data analysis at different scales.However,the computational complexity of the tensor decomposition is relatively high,and the practical application has high requirements for the calculation efficiency.In this paper,the tensor decomposition and reconstruction methods based on the Graphics Processing Unit(GPU)are studied for the efficiency of tensor decomposition.For GPU-based tensor decomposition and reconstruction problems,this paper conducts research work from GPU performance model and GPU-based Tucker decomposition and reconstruction.The main contents and contributions are as follows:1.A GPU performance analysis model was proposed.Aiming at the performance evaluation of GPU programs,a quantified performance analysis model can predict the performance of application migrated to GPUs and evaluate bottlenecks,and help to understand GPU programming model and performance optimization.Based on this,this paper proposes a GPU performance analysis model,which takes into account not only the overhead incurred by GPU execution,but also the quantitative analysis of the program by modeling the thread calculation instructions and memory access instructions from the perspective of the instruction pipeline,and understand program performance characteristics.Through simulation analysis,the proposed model can predict the performance of GPU program with high accuracy.2.A GPU-based Tucker decomposition and reconstruction algorithm is proposed.For Tucker decomposition and reconstruction algorithm,the computation complexity is too high,this paper presents a GPU-based tensor decomposition and reconstruction parallelization algorithm.By analyzing the performance of the algorithm,the most time-consuming series of tensor and matrix multiplication is paralledized based on GPU,and intermediate data reuse is used to optimizes data transmission cost between CPU and GPU,accelerate Tucker decomposition and reconstruction.Through simulation test,the proposed method achieves several times speedup.3.The visualization application of large-scale 3D seismic data based on GPU tensor decomposition and reconstruction is realized.There are problems with the existing visualization methods that can only provide specific resolutions,a visualization method based on continuous multi-resolution is proposed,and the parallelization research results of tensor decomposition and reconstruction are applied in the visualization system.In summary,the GPU performance analysis model helps understand the GPU programming model and performance optimization,GPU-based tensor decomposition and reconstruction proposed and implemented can improve computational efficiency in the visualization of large-scale 3D seismic data.