Exploiting motion estimation resilience to approximated metrics on SIMD-capable general processors: From Atom to Nehalem

In the past, efforts to speed up motion estimation for video encoding were directed at finding better predictive search algorithms. Now, they are directed toward the shrewd exploitation of the machine's advanced architectural features such as multimedia extensions, especially for the computation of the error metric which is known to be expensive. In this paper, we extend previous work by further exploring efficient implementation of approximate fast metrics for motion estimation. We show that the proposed metrics can be implemented using SIMD instructions to yield impressive speed-ups, up to 12:1 relative to non-vectorized but otherwise optimized C code, while sacrificing less than 0.1 dB on image quality.