Evaluating color instruction set extension for real-time vector quantization

Abstract

Vector quantization (VQ) is widely used for color image and video compression. However, its high computational overhead prohibits many applications in real-time systems. This paper presents a novel method to accelerate full-search VQ algorithm by adding quantized color pack extension (QCPX) instruction set architecture (ISA). QCPX not only supports a packed 16-bit YCbCr data format but also obtains performance and code density improvements through three-color pixels in parallel in a 16-bit width. To measure execution performance of the QCPX instruction set architecture (ISA), it is evaluated in a SIMD pixel array platform developed at Georgia Tech. In addition, by varying the grain size (pixel per processing element, PPE), this study can fully measure the impact of QCPX in the presence of different levels of data parallelism. Simulation results indicate that QCPX version achieves speedups from 27% to 297% over non-QCPX with the most impressive improvements >200 % occurring above the communication-bound 16 PPE granularity. QCPX also reduces average PE idle cycles by 45%. QCPX can be incorporated in range of architectures from current ILP processors to future massively data parallel machines