Parallel Algorithm Research And Realization Of Linear Dimensionality Reduction For Hyperspectral Image On CPU/GPU

How to efficiently process hyperspectral information has become the research focus of remote sensing area. Dimensionality reduction of hyperspectral image is the key step of the following analysis and identification of landscape. Dimensionality reduction is a typical computing intensive and memory access intensive process with high complexity. Realizing data reduction in serial mode is impossible to satisfy the real-time need of many applications, such as military and agriculture. Heterogeneous system with general CPU and special GPU can satisfy different needs for computing resources, and is one of the most promising developments of future high performance computer architecture. How to effectively utilize the computing power of the heterogeneous systems is the key of this paper. At present and in the future, CPU + GPU heterogeneous system, which can satisfy the demand of application of highly intensive computing, is one of the mainstream development direction of high performance computer architecture. How to make full use of the computing ability of the CPU/GPU heterogeneous system to improve the processing speed of linear dimension reduction of hyperspectral remote image is the key of this study.There are three parallel levels in the CPU/GPU heterogeneous system, such as processes between nodes, threads in a node and threads in the GPU. The parallel programming models of MPI, openMP, CUDA are chosed in this paper. PCA and Fast ICA, which are typical linear dimension reduction algorithms, were selected. we researched the strategies of parallel and optimization of different accelerations, and came up with some parallel algorithms of linear dimension reduction. In this paper, we have done the following works and innovations:(1)Further study of 3 current parallel modes of CPU/GPU heterogeneous system was done. Multiple CPUs, GPUs and nodes are the characteristics of heterogeneous systems. the technologies of MPI, openMP and CUDA must be mastered. In this paper, we studied the propramming models and the optimization strategies of signal and hybrid parallel.(2)A series of parallel PCA algorithms were come up on the CPU/GPU heterogeneous system. On the basis of analyzing accelerate hots of PCA algorithm, we put forward the task assignment for different parallel levels, and came up with the parallel and optimization strategies of the hot steps of covariance matrix calculation, PCA transform and hyperspectral data I/O. There are 6 kinds of multi-parallel PCA algorithms proposed respectively based on MPI, openMP, CUDA, MPI+CUDA, openMP+CUDA and MPI+openMP+CUDA. The experimental results show that the 6 kinds of parallel algorithms all obtained the remarkable performance improvement, and the triple parallel PCA algorithm of MPI+openMP+CUDA get 145 times speed-up(without I/O) and 128 times of sum speed-up(include I/O).(3)A series of parallel FastICA algorithms were proposed based on the CPU/GPU heterogeneous system. With analyzing accelerate hots of FastICA algorithm, we come up with the parallel and optimization strategies of the hot steps of white processing, the ICA iterations and IC transformation. An innovation of parallel design mode of “concrete – abstract- concrete” was put forward to complex calculation of ICA iterations in particular. We have come up with 6 kinds of multi-parallel Fast ICA algorithms. Experiments showed that the acceleration of the 6 kinds of parallel algorithm is ideal. 169 times speedup(without I/O) and 159 times speedup(include I/O) were taken by the triple parallel FastICA algorithm of MPI+openMP+CUDA.