A Spatial Clustering Approach for Reduced Memory Traffic in Motion Estimation using Intelreg; Wireless MMX~(TM) Technology

摘要

Video compression is one of the more demanding application domains targeting modern desktop and mobile platforms. User demands for visual communications are driving increases in resolution, framerate, and perceptual quality. The resulting processing requirements are inspiring both new hardware architectures and more efficient algorithms. This paper introduces a spatial clustering approach for the implementation of fixed pattern fast motion search with low memory accesses and computation time using Wireless MMX Technology. Intel® Wireless MMX™ technology is a 64-bit Single Instruction Multiple Data, (SIMD), coprocessor for the Intel® XScale® microarchitecture. The technology enables a fully programmable solution for implementation of motion estimation algorithms. Specific architectural features are introduced which allow the reuse of search region data through load amortization and alignment control in a manner previously seen only with VLSI architectures. Our approach allows reference pixel data to be amortized in two dimensions so redundant memory accesses can be reduced or eliminated completely as a search algorithm proceeds.