CUDA-Accelerated ODETLAP : a parallel lossy compression implementation for multidimensional data

Abstract

This thesis presents an efficient ODETLAP implementation for the compression of gridded, multidimensional data. As vast quantities of data are collected for geographic information systems, compression allows for the storage and transmission of larger data sets. Techniques that utilize autocorrelation in all data dimensions allow for greater levels of compression but are more computationally intensive. ODETLAP, Overdetermined Laplacian Approximation, uses a subset of points from the original data set to accurately reconstruct the data. As it expands into higher dimensions, ODETLAP is capable of using relationships in data across multiple dimensions. Various parallelization techniques are used to improve computation time, utilizing CUDA for general purpose programming on a graphics processing unit (GPGPU). An efficient ODETLAP implementation was created directly in GPU memory, successfully avoiding the overhead associated with the transfer of data between GPU memory and main memory.