Abstract
Context-based Adaptive Binary Arithmetic Coding (CABAC) in H.264/AVC can achieve high compression ratio at the expense of high computational complexity. We have previously proposed a fully hardwired CABAC decoder that supports real-time QFHD (4x1080HD) decoding at the maximum bit rate of 80 Mbps. For higher end applications, we analyze the bin distribution of each Syntax Element (SE) type and the performance of our previous work. The analysis results show that mvd SEs account for significant amount of bins and Get-Neighbor (GN) process degrades the utilization of Arithmetic Engine (AE). Therefore, we propose three methods to speed up mvd decoding and increase AE utilization. For mvd SEs, we use a Two-Bin Arithmetic Engine (TBAE) to decode two mvd bins per cycle. To increase AE utilization, we reduce the cycle-count of the GN process by balancing its critical path. In addition, we propose a prediction method to perform AE and GN in parallel. Experimental results show that our new CABAC decoder can achieve 45% throughput (in bins per second) improvement and is capable of real-time decoding QFHD video at the maximum bit rate of 221 Mbps when running at 238 MHz.