Context-Base Computational Value Prediction with Value Compression

Context-Base Computational Value Prediction with Value Compression
复制标题

DOI:
--
复制
发表时间:
2018
期刊:
--
影响因子:
--
通讯作者:
Y. Ishii;Arm;Y. Ishii
Y. Ishii;Arm;Y. Ishii
中科院分区:
其他
文献类型:
--
作者:
Y. Ishii;Arm;Y. Ishii

文献摘要

被引文献

相似文献

在本文中,我们提出了基于上下文的计算价值预测和利用价值局部性的激进成本降低技术。基于上下文的计算值预测跟踪相应上下文的基本值及其跨步,以覆盖依赖于上下文的增量模式。我们还提出了一种使用值压缩缓存的激进值压缩方案。该缓存跟踪跟踪值的高阶位。为了使存储效率最大化,只有在预测值具有一定置信度后才能缓存数据。我们将这些建议和动态覆盖精度控制应用于TAGE启发值预测器CBC-VTAGE。在8KB的存储预算下,我们优化的CBC-VTAGE与没有值预测的处理器相比,速度提高了17.1%。分布式135走线的IPC几何系数为3.76。
In this paper, we propose context-base computation value prediction and aggressive cost reduction techniques exploiting value locality. Context-base computational value prediction tracks both base value and its stride for corresponding context to cover context-dependent delta patterns. We also propose an aggressive value compression scheme using value compression cache. This cache tracks higher order bits of tracking values. To maximize the storage efficiency, the data can be cached only after the predicting value have a certain confidence. We applied these proposals and dynamic coverage accuracy control on TAGE inspired value predictor, CBC-VTAGE. With 8KB storage budget, our optimized CBC-VTAGE achieves 17.1% speedup compared with processor without value prediction. The geomean of IPC for distributed 135 traces is 3.76.