An effective implementation of vector quantization encoder on commodity graphics hardware

This paper presents an effective implementation of vector quantization (VQ) encoder on a commodity personal computer (PC). A programmable graphics processing unit (GPU) is used for finding the nearest codeword to each data vector, so-called nearest neighbor search, which dominates the most execution time of the VQ encoding process. In our implementation, the nearest neighbor search is bound to GPU's programmable rendering pipeline so that the GPU can well exhibit its parallel-processing ability. Experimental results show that the single-instruction multiple-data nature of the nearest neighbor search allows the GPU to efficiently execute the vector quantization encoding process. Comparing with vector quantization encoder fully implemented on a general-purpose microprocessor, our implementation can achieve better computational efficiency for the VQ encoding process with a large codebook.