Parallel Optimization On GPUs For Intra Coding Algorithms Of High Efficiency Video Coding

Intra coding is one of the key parts of H.265, which is the new generation high efficiency video coding(HEVC) standard. With advanced prediction techniques, intra coding can greatly improve the video coding efficiency. However, the coding process become more complicated, which will bring challenges for real-time coding. Due to the powerful multi-level parallelism support, the emerging Graphics Processing Units(GPUs) provide a platform for the real-time realization of intra coding. Therefore, how to efficiently mine the parallelism of intra coding and combine it with GPUs is very important for the practical application of H.265 video coding.Based on the analysis of intra coding's data level parallelism and thread level parallelism, a GPU parallel optimization solution for H.265 intra coding is proposed. To efficiently mine the parallelism of intra coding and speed up the process of intra coding, a level parallelism relationship model and an Open CL(Open Computing Language) multiple level parallel method are proposed. Firstly, the main process of H.265 intra coding are being analyzed, and the data dependency of five directions is established on the basis of the reference sample position required by multiple prediction. Then the level parallelism relationship model is proposed according to the coding process of multiple unit level. Secondly, the high-parallelism characteristic of GPUs are utilized to design the corresponding parallel strategies. Then the implementation details like ND-range(N-Dimensional Range) division are optimized combining the characteristics of Open CL coding, and taking advantage of the analytical bitrate estimation method, the Open CL multiple level parallel strategy is also proposed. Combining the parallelism relationship model and the Open CL multiple level parallel strategy, the GPU parallel optimization solution for H.265 intra coding is developed.Experimental results suggest that after applying the proposed GPU optimization solution, the speed of intra coding part could be at most 0.89 times faster than the HM(HEVC Test Model) encoder, with a video quality lose in PSNR(Peak Signal to Noise Ratio) within 0.75 d B. And if more video quality lose in PSNR(1.21 d B) is allowed, the speed may be at most 1.42 times faster. The proposed solution could efficiently improve the speed of encoding with an acceptable video quality lose.