Optimizing Motion Estimation with an ReRAM-Based PIM Architecture

Motion estimation (ME) is an HEVC process for determining motion vectors that describe the blocks transformation direction from one frame to a future adjacent frame in a video sequence. ME is a memory and computationally intensive process which consumes more than 50% of the total running time of HEVC. To remedy the memory and computation challenges, in this paper, we present ReME, a highly paralleled Processing-In-Memory accelerator for ME based on ReRAM. In ReME, the space of ReRAM is separated into storage engine and ME processing engine. The storage engine acts as the conventional memory to store video frames and intermediate data while the processing engine is for ME computation. Each ME processing engine in ReME consists of a SAD (Sum of Absolute Differences) model, an interpolation model, and a SATD (Sum of Absolute Transformed Difference) model that transfer ME functions into ReRAM-based logic analog computation units. ReME further cooperates these basic computation units to perform ME processes in a highly parallel manner. Simulation results show that the proposed ReME accelerator significantly outperforms other implementations with time consuming and energy saving.