Parallel H.264 encoding algorithms for general purpose computing GPUs

Abstract

As of 2000, the direction of processor evolution has turned from increasing the clock frequency to increasing the number of cores on a chip. Also, new architectural features such as scratchpad memory and reconfigurable processors are introduced. One type of processors that is leading the frontier is the GPU. Having 10s of cores and featuring scratchpad memory and ability to reconfigure resources allocated to a core depending on the application, GPUs can be an excellent prototype of future manycore processors. It is important for the current compute-intensive applications to adapt themselves to such highly parallel architectures. Video encoding applications are one of the most compute-intensive applications that are used very broadly and frequently nowadays. Video encoding algorithms have been developed enough to utilize the full potential of current multicore architectures. However, the current video coding algorithms exhibit insufficient parallelism to fully utilize highly parallel architectures such as GPUs. In this paper, highly parallel algorithms for H.264 video encoding are introduced. Different from the current parallel encoding approaches, the new algorithms tries to minimize the data dependencies due to hint information, thus providing more parallelism while leaving the coding efficiency almost unchanged. Experiment shows that the GPU implementation running on GTX260 performs 3.95x faster than x264 open source encoder running on a quadcore Intel Xeon 2.4 GHz processor, while leaving the coding efficiency almost unchanged for sample video sequences with regular motion.