An Efficient Pipelined Architecture for H.264/AVC Intra Frame Processing

An implementation of intra prediction with transform for H.264/AVC

“…If we take such intra prediction and transform logic for an H.264/AVC encoder with inter prediction function, the latency is increased because the transform logic must be used for chroma transform once the luma prediction is finished. Most of such works [4,5,6,7] used modified intra prediction algorithms to accelerate intra prediction at the cost of intra prediction quality. The transform logic is shared in [7], but the cycles/MB is very low even though I4 and I16 predictions are interleaved due to their fast prediction algorithm.…”

“…Most of such works [4,5,6,7] used modified intra prediction algorithms to accelerate intra prediction at the cost of intra prediction quality. The transform logic is shared in [7], but the cycles/MB is very low even though I4 and I16 predictions are interleaved due to their fast prediction algorithm. The design presented in [10] shows excellent cycles/MB, but the work only considered I4 prediction.…”

“…Lin et al presented an efficient mode decision algorithm that finds the most probable mode with less image quality degradation [6]. Jin et al presented a highly pipelined architecture for H.264/AVC intra prediction [7] and Suh et al presented an efficient architecture for H.264 with full-mode search particularly focused on reducing the execution cycles for intra 16×16 (I16) predictions [8]. Lee et al proposed rearranging the encoding order of intra 4×4 (I4) predictions to increase the throughput of the pipelined process of intra prediction and transform [9].…”

Fast H.264 intra prediction for network video processing

“…The decision-algorithm, responsible for selecting that prediction mode which is believed to result in the smallest bitstream, is not specified by the standard [2]. Though more complex algorithms have been developed in the past [5,7], for real-time implementation on dedicated hardware the brute force approach followed by summing of the residuals (Sum of Absolute Differences, SAD) to make the decision is popular [4].…”

“…In [4] an overview is presented of previously developed techniques to speed up the intra prediction process, by pipelining calculations, exploiting parallelism, changing the order in which blocks are processed [5] and even skipping prediction modes [5]. Furthermore, [4] presents a pipeline which supports the same modes as in this work, though no performance is mentioned.…”