Priority-Based Heading One Detector in H.264/AVC Decoding

Ke Xu, Chiu‐Sing Choy, Cheong-Fat Chan, Kong‐Pang PunEURASIP 2007 · DOI

TL;DR

This paper introduces a novel, power-efficient detector for identifying the "heading one" in H.264/AVC video decoding. By exploiting the statistical distribution of these "heading ones," their proposed detector significantly reduces power consumption (over 3x) compared to existing methods, without compromising performance or increasing hardware cost.

What Problem Does This Paper Address?

Modern video codecs like H.264/AVC offer significant improvements in compression efficiency, but at the cost of greatly increased computational complexity and power consumption. This is particularly problematic for portable devices where power is a critical design constraint. A key component in H.264/AVC decoding, specifically for entropy coding methods like Exp-Golomb and CAVLC, is the "heading one detector." This detector's job is to find the position of the first '1' bit within a given codeword. Existing detectors, like the one described in [4], treat all bits equally, leading to unnecessary power usage even when the "heading one" is found early in the bitstream. This paper addresses the need for a more power-efficient heading one detector that intelligently adapts to the statistical nature of the data.

Methods

The core innovation lies in a "priority-based nonuniform partition" approach. Instead of dividing the 16-bit input uniformly, the authors divide it into three unequal parts with varying priorities.

  • Statistical Analysis: The paper first analyzes the statistical distribution of heading one positions across various video sequences (intra- and inter-coded frames). They observe that the heading one is overwhelmingly likely to appear in the first few bits (e.g., position 0 or 1).
  • Nonuniform Partitioning: Based on this statistical insight, the 16-bit input is partitioned into three segments:
  • Dec2: Handles the first two bits (highest priority, always active).
  • Dec4: Handles the next four bits (medium priority, active if Dec2 fails).
  • Dec10: Handles the remaining ten bits (lowest priority, active only if Dec2 and Dec4 fail).
  • Selective Enabling: This hierarchical structure allows lower-priority detectors (Dec4 and Dec10) to be disabled when a heading one is found by a higher-priority detector, thus saving power.
  • Priority Encoder: A final priority encoder combines the outputs from the active sub-detectors to determine the heading one position.

This contrasts with prior work [4] that uses uniform partitioning and keeps all sub-detectors and priority encoders active simultaneously, regardless of where the heading one is found.

Datasets

The paper uses five popular test videos: "container," "foreman," "akiyo," "news," and "carphone." These videos are encoded using the JM software (version 9.4) at QCIF resolution and 30 fps, with quantization parameters set to 22, 25, 28, and 32. This allows for the statistical analysis of heading one positions in both intra- and inter-coded frames.

Key Insights and Ideas

  • Statistical Redundancy is Key: The most significant insight is that the "heading one" in H.264/AVC codewords is not uniformly distributed. It's highly concentrated in the initial bits, which can be exploited for power savings.
  • Data-Driven Design: The proposed detector is "data-driven," meaning its operation and power consumption are directly influenced by the statistical characteristics of the input data, rather than a fixed, uniform processing approach.
  • Selective Activation Saves Power: By disabling less critical processing units when a result is found early, significant power savings are achieved. This is a fundamental principle of efficient hardware design.
  • Trade-off Management: The design acknowledges a slight throughput degradation in extreme cases (heading one at a very high position) but argues that the overall power savings and maintained performance are well worth this trade-off for most applications.

Key Results

  • Power Reduction: The proposed priority-based detector achieves over 3 times power reduction compared to a conventional detector. Implementation analysis shows a power reduction of 3.38 times for intra-frames and 3.21 times for inter-frames.
  • Performance: The design maintains performance and can achieve a maximum frequency of 200 MHz, which is sufficient for real-time decoding.
  • Area: The circuit size is comparable to, and slightly smaller than, the design in [4], due to the reduction in the number of priority encoders.
  • Throughput: While there's a slight throughput penalty in rare cases, the overall throughput is comparable to existing methods.

Limitations and Future Work

  • Throughput Degradation: The authors acknowledge that in rare cases where the heading one appears at a very high bit position (e.g., 6 or higher), the sequential activation of dec2, dec4, and dec10 might lead to a slight throughput degradation.
  • Generalizability: The paper suggests that the proposed technique, based on exploiting statistical codeword distribution, can be applied to other entropy coding schemes beyond H.264/AVC, such as those in MPEG-2, where short codewords also occur more frequently. This implies a broader applicability for the core concept.

Summary updated

BibTeX
@inproceedings{xu2007priority,
  author = {Ke Xu and Chiu‐Sing Choy and Cheong-Fat Chan and Kong‐Pang Pun},
  title = {{Priority-Based Heading One Detector in H.264/AVC Decoding}},
  booktitle = {EURASIP},
  year = {2007},
  doi = {10.1186/1687-3963-2007-060834},
}
Figure