Parallel architectures for entropy coding in a dual-standard ultra-HD video encoder

[摘要] The mismatch between the rapid increase in resolution requirements and the slower increase in energy capacity demand more aggressive low-power circuit design techniques to maintain battery life of hand-held multimedia devices. As the operating voltage is lowered to reduce power consumption, the maximum operating frequency of the system must also decrease while the performance requirements remain constant. To meet these performance constraints imposed by the high resolution and complex functionality of video processing systems, novel techniques for increasing throughput are explored. In particular, the entropy coding functional block faces the most stringent requirements to deliver the necessary throughput due to its highly serial nature, especially to sustain real-time encoding. This thesis proposes parallel architectures for high-performance entropy coding for high-resolution, dual-standard video encoding. To demonstrate the most aggressive techniques for achieving standard reconfigurability, two markedly different video compression standards (H.264/AVC and VC-1) are supported. Specifically, the entropy coder must process data generated from a quad full-HD (4096x2160 pixels per frame, the equivalent of four full-HD frames) video at a frame rate of 30 frames per second and perform lossless compression to generate an output bitstream. This block will be integrated into a dual-standard video encoder chip targeted for operation at 0.6V, which will be fabricated following the completion of this thesis. Parallelism, as well as other techniques applied at the syntax element or bit level, are used to achieve the overall throughput requirements. Three frames of video data are processed in parallel at the system level, and varying degrees of parallelism are employed within the entropy coding block for each standard. The VC-1 entropy encoder block encodes 735M symbols per second with a gate count of 136.6K and power consumption of 304.5 pW, and the H.264 block encodes 4.97G binary symbols per second through three-frame parallelism and a 6-bin cascaded pipelining architecture with a critical path delay of 20.05 ns.

[发布机构] Massachusetts Institute of Technology