CUVLE: Variable-Length Encoding on CUDA

Abstract

Data compression is the process of representing information in a compact form, in order to reduce the storage requirements and, hence, communication bandwidth. It has been one of the critical enabling technologies for the ongoing digital multimedia revolution for decades. In the variable-length encoding (VLE) compression method, most frequently occurring symbols are replaced by codes with shorter lengths. As it is a common strategy in many compression applications, efficient parallel implementations of VLE are very desirable. In this paper we present CUVLE, a GPU implementation of VLE on CUDA. Our approach is on average more than 20 and 2 times faster than the corresponding CPU serial implementation and the only known state-of-the-art GPU implementation, respectively.