Research Of H.264Parallel Compression Algorithm With CUDA

With the development of network and3G technology, it promotes theapplications and research of video compression standard with high compressionefficiency and better network robustness in video communications. However, thecomputationally intensive and data-intensive of video compression coding standardsare still constraints in the application, such as H.264/AVC. NVIDIA’s CUDA showsthe high-performance computing of GPU and can apply its power to scientificcomputing, graphics and video processing which has high computing demands in dataprocessing. Based on CPU and GPU heterogeneous parallel system, the research ofapplying it to video encoding technology and the key algorithms of video compressioncoding will be significantly important to improve multimedia application technologyand high-performance computing technology.This paper took the CUDA-based parallel computing and motion estimationalgorithm as the main object of study and propsed a CUDA-based parallel computingmodel of motion estimation and optimization model. The results of this studyinclude:Firstly, it analyzed the parallel computing model of CPU/GPU heterogeneoussystems and proposed a global motion estimation parallel computing model. At thebeginning, it studied the characteristics of global motion estimation algorithm andcompleted motion estimation matching by iteration matrix of multiple threads onCUDA. Based on asynchronous mechanism of CUDA architecture, it can hide partdelay which was caused by kernel function. This model of motion estimation hasmade significant acceleration efficient.Secondly, it proposed a fast searching motion estimation algorithm that wassuitable for parallel computing model based on CUDA. It took predicting macro-blockas the smallest unit of parallel computation in the block and balanced the computationamong warps; each warp processed the matching work between one or morepredicting macro-block and current block. This model, divided parallel computingtasks with warps, achieved data partition model based on CUDA and was moresuitable for variety quick search models.Thirdly, we proposed a CUDA-based optimization model. The model analyzesvarious hardware resources on GPU and its impact on performance of parallelprogram. Then, it tested parallel program with CUDA profiler: comparing and analyzing the relationship of different number of warps, GPU resource occupancy, andparallel program performance.