Research And Optimization Of Adaptive Interpolation Filter In H.265 With CUDA

H.265 is a Video Coding Standard, being planned by ITU-T VCEG (video coding experts group), to provide better video coding methods for audio and video applications such as conversational and non-conversational service. In H.265, in addition to some traditional coding methods, some new technologies are adopted to balance the relationship among image quality, coding delay, and complexity.Adaptive Interpolation Filter is one of the new techniques used in H.265. Compared with the 6-tap bilinear filter used in H.264, which results in aliasing noise to some extent and therefore affect the accuracy of motion estimation especially for sub-pixel based motion vector, AIF is used to reduce the aliasing noise since it is a Wiener filter and its coefficients are chosen based on least square error constraint.Graphic Process Unit (GPU) is a specialized microprocessor that offloads and accelerates graphics rendering from the central processor, modern GPU provide incredible resource for graphic processing and general purpose application. For example, GPU's float operation performance today can reach 54GLOPS, which is nearly ten times of Intel Core2 Duo (about 5.6 GFLOPS). In order to make full use of the computation ability in GPU, more and more efforts have been made to offload computation task from CPU to GPU. With sweeping research and application in exploiting computing resources of SIMD (Single Instruction Multi-Data) stream processor in GPU, it becomes an effective coprocessor to CPU and derives highly cost performance.This paper starts with a general introduction to the key technologies used in H.264 and new technologies adopted by H.265. Then, it focuses on the principle and implementation of AIF technologies. The hardware and software framework related with CUDA (Compute Unified Device Architecture) are also addressed to illustrate the outstanding characteristics of CUDA especially in general computation based applications. The next part of this paper reveals all details to speed up AIF operation by CUDA. Performance assessment is also undertaken using KTA (Key Technical Area) reference software to demonstrate the advantages of CUDA for image filtering. Simulation results show that CUDA based parallel optimization can improve the coding speed of AIF by a factor of 2 as compared with traditional implementation solution.