An Efficient Design of H.264 Inter Interpolator with Bandwidth Optimization

Abstract

In this paper, a highly efficient inter-interpolation architecture for the H.264/AVC standard is proposed. Since the placement order of frame pixels in the memory is either row-wise or column-wise which may not be suitable for the sample prediction in particular direction, this paper proposes a novel interpolator design which can dynamically configure the data-path for different predicted modes to perform proper computation schedules suitable for the nature input order of reference samples. The proposed design methodology not only can avoid the additional data transposition buffer, but most importantly the data transfer time spent to fetch the reference samples can be overlapped with the data computation time. Furthermore, by decomposing the chroma interpolation into a series of shift and addition operations, both luma and chroma interpolations can be integrated within the same module. In addition to the data-path design, this paper also proposes a new data-reuse buffer design based on a two-dimensional cache architecture to explore the possible data reuse among the inter and intra partitions. This design can be easily integrated with the H.264 interpolator to reduce the enormous demand of memory access. Our experimental results shows that our saving of memory bandwidth can be 23% more than what the best design can achieve by exploring the intra-partition data reuse only. The proposed design methodology has been implemented, and the result shows that the proposed interpolation architecture is the most compact design among the literatures which can perform the real-time H.264 video decoding with resolution up to 1920×1088 high-definition television standard. The proposed interpolator can be applied to the dedicated H.264 hardware codec design for various consumer devices.