摘要
子字并行能够充分利用多媒体算法的数据精度小、内部循环处理形式规则的特点,是加速多媒体处理的有效方式。然而,如何充分挖掘多媒体应用中的子字并行仍然是一个难题。本文说明传统的并行技术可以有效地开发循环中的子字并行性,同时提出一种基于代价子图的子字并行指令自动识别的方法。与其他方法相比,该方法利用代价模型对子字并行指令选择进行定量评估。本文在TTA体系结构框架下实现了这一方法。实验结果表明,该方法可以充分地提取循环中的子字并行性。
Subword parallelism can fully utilize the characteristics of multimedia algorithms, and it is an effective way to accelerate multimedia processing. However it is hard to mine the subwords in multimedia applications. This paper shows that the traditional parallelization techniques can be used to exploit subword parallelism, and also proposes a novel method to extract subword parallelism based on the cost subgraph. We evaluate the effectiveness of our methods for a number of benchmarks on the TTA framework. The results reveal that this method can significantly obtain the available subword parallelism in the loop.
出处
《计算机工程与科学》
CSCD
2008年第9期141-144,150,共5页
Computer Engineering & Science
基金
国家自然科学基金资助项目(60173040)
关键词
子字并行
指令选择
代价子图
subword parallelism
instruction selection
cost subgraph
