Energy analisys of motion estimation memory transference on embedded processors

This paper presents a memory-transference analysis to a parallel Motion Estimation (ME) algorithms for current embedded processors, that usually are composed by a CPU and GPU with OpenCL parallel-programming support. However, in this scope, the CPU and GPU memories are different, thus being necessary a memory transference data between then. This paper introduces the main concepts of the ME, discusses its related problems and proposes different approaches for CPU and GPU memory transference. Three different approaches for reference frame transference was evaluated and tested using three different ME algorithms. The experiments evaluated the time performance and the energy consumption of all tests considering each proposed memory transference approaches. The results indicate that the best solution of memory transference is using the Full Frame approach, where each reference frame was transferred to the GPU memory for every new current frame.