High-Throughput Variable-to-Fixed Entropy Codec Using Selective, Stochastic Code Forests

Abstract

Efficient high-throughput (HT) compression algorithms are paramount to meet the stringent constraints of present and upcoming data storage, processing, and transmission systems. In particular, latency, bandwidth and energy requirements are critical for those systems. Most HT codecs are designed to maximize compression speed, and secondarily to minimize compressed lengths. On the other hand, decompression speed is often equally or more critical than compression speed, especially in scenarios where decompression is performed multiple times and/or at critical parts of a system. In this work, an algorithm to design variable-to-fixed (VF) codes is proposed that prioritizes decompression speed. Stationary Markov analysis is employed to generate multiple, jointly optimized codes (denoted code forests). Their average compression efficiency is on par with the state of the art in VF codes, e.g., within 1% of Yamamoto et al.'s algorithm. The proposed code forest structure enables the implementation of highly efficient codecs, with decompression speeds 3.8 times faster than other state-of-the-art HT entropy codecs with equal or better compression ratios for natural data sources. Compared to these HT codecs, the proposed forests yields similar compression efficiency and speeds. INDEX TERMS Data compression, high-throughput entropy coding, variable-to-fixed codes.

Streaming-Capable High-Performance Architecture of Learned Image Compression Codecs

“…Even with our architecture, learned image encoding and decoding is still a highly GPU-bound algorithm on most platforms. Therefore, as future work, by combining our method with other research to optimize the neural model, even higher performance could be achieved; on the other hand, the introduction of higher performance entropy coders such as [16] and [19] could potentially further reduce latency on desktop platforms. Especially for context-based models, combining our architecture with works such as [17] and [18] enables more parallel execution and higher throughput to be achieved.…”

“…On the other hand, works like [16] and [19] showed different approaches in implementing high-performance entropy coders. [17] and [18] demonstrated different methods enabling parallel entropy parameter calculations in context models, significantly boosting the serial masked CNN performance bottleneck.…”

“…Unlike previous works, our method works on most existing learned image codecs without either new neural model or entropy coder architectures. Our approach also comes with the capability to be further optimized when combined with the above model optimizations such as [7], [14], and [6], or highperformance entropy coders such as [16] and [19].…”