Implementation of H.264 intra-frame encoding on clustered stream architectures

Most videos are encoded by using H.264 standard to obtain a high compression ratio. However, the high computation complexity and strong data dependency make the real-time encoding difficult. Since stream architectures have abundant arithmetic units and are flexible programmable, they are more suitable for media applications than ASICs and DSPs. This paper presents the real-time implementation of H.264 intra-frame encoding on the clustered stream architecture. Optimizations at both stream level and kernel level are considered to improve the utilization of arithmetic units and then the encoding performance. At stream level, the parallelisms between blocks, macroblocks, and slices are exploited. At kernel level, kernel partition and kernel fusion are combined to get optimized kernels. The experimental results demonstrate that the proposed optimizing techniques enhance the performance significantly. The intra-frame encoding of 1080p video could achieve encoding rate of 50.2 fps.