Implementation of Block Matching Algorithms on CUDA

Abstract

With the increasing popularity of technologies such as Internet streaming video and video conferenting, video compression has became an essential component of broadcast and en- tertainment media. Motion Estimation (ME) and compensation techniques, which can eliminate temporal redundancy between adjacent frames e ectively, have been widely ap- plied to popular video compression coding standards such as MPEG4 AVC/H.264. The use of a high-level parallelism of GPU architecture for the high de nition images and video processing give high performance and high speed of data processing. Due to its object based nature, MPEG4 AVC is good candidate for parallelism. Among all steps of encoder, motion estimation is very time consuming step. The wide range of memories o ered by the GPU architecture has been exploited to a large extent and a signi cant speedup of motion estimation on the GPU as compared to the serial implementation. The use of texture memory for accessing the periphery of the frames prevents the out of bound references and contributes to the speedup achieved. The use of shared memory for processing current frame and reference frame exploit signi cant speed for temporal redundancy. The parallel versions of Full search and Diamond search are implemented on Tesla C2070 on Ubuntu OS using CUDA4.0 technology. The results are presented along with methodology used for parallelization of code for NVIDIA GPU. We also ob- tain increased performance by reordering accesses to o -chip global memory to combine requests to the same or contiguous memory locations and apply classical optimizations to reduce the number of executed operations.