A study on a method of effective memory utilization on GPU applied for neighboring filter on image processing

Abstract

In this paper, the methods of implementing neighboring filters on newly supplied Graphics Processing Unit (GPU) are described. In general, neighboring filters are always utilized in image processing. Mainly in consideration of memory accesses, four methods implementing neighboring filtering are proposed and compared. The experimental result shows that one of the proposed methods (called "4X-block") at the block size of 16 is the fastest among them, when loading and processing data in shared memory in GPU. It is also shown that this method is about 1.45X faster than the basic method implemented on GPU. © 2012 Springer-Verlag GmbH.