| 承书尧* **,郝一帆*,杜子东*.自动流水设计中的门级依赖分析算法[J].高技术通讯(中文),2025,35(9):943~950 |
| 自动流水设计中的门级依赖分析算法 |
| Gate-level dependency analysis in automated pipeline design |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 09. 003 |
| 中文关键词: 人工智能； 电子设计自动化； 机器学习 |
| 英文关键词: artificial intelligence, electronic design automation, machine learning |
| 基金项目: |
| 作者 | 单位 | | 承书尧* ** | (*中国科学院计算技术研究所处理器芯片全国重点实验室北京 100190)
(**中国科学院大学北京 100049) | | 郝一帆* | | | 杜子东* | |
|
| 摘要点击次数: 26 |
| 全文下载次数: 23 |
| 中文摘要: |
| 自动流水设计是一种经典的电子设计自动化（electronic design automation，EDA）技术，在处理器电路设计、高层次综合领域被广泛研究。其核心挑战在于处理“写后读”数据依赖。现有方法中的“写后读”依赖分析基于高层次人工规则，缺乏细粒度门级电路信息，导致设计过于保守，使得自动流水设计的并行效率和面积开销均无法与人类设计相比。本文提出一种数据驱动的门级依赖分析算法，其核心观察是流水设计的“写后读”依赖可以由电路模拟执行得到的输入-输出采样进行更精确的细粒度描述。基于此，算法形式化给出自动流水设计中的功能约束和性能约束，对“写后读”依赖分析问题进行更细粒度的建模，并构建基于二元推测图拓展的机器学习算法进行求解。本文在标准电路数据集上进行实验，结果表明，自动流水设计的并行性能和人类设计相当，综合性能超越了现有的自动流水设计方法，吞吐量平均性能提高19.0%。 |
| 英文摘要: |
| Automatic pipeline design is a classic electronic design automation（EDA） technology widely studied in the processor circuit design and high-level synthesis. The main challenge is to handle the‘Read after Write’ data dependencies. The existing methods for ‘Read after Write’ dependency analysis is based on high-level design rules without fine-grained gate level circuit information, making the parallel efficiency and area cost of automatic pipeline design not comparable with human design. This article proposes a data-driven gate level dependency analysis algorithm. The core observation is that the ‘Read after Write’ dependency of pipeline design can be more accurately described at a fine-grained level through input-output sampling obtained through circuit simulation. Based on this, the algorithm formalizes the functional and performance constraints in automatic pipeline design, models the ‘Read after Write’ dependency analysis problem at a finer granularity, and constructs a machine learning algorithm based on binary inference graph extension for solution. This article conducts experiments on standard circuit datasets, and the results show that the throughput performance of automatic pipeline design is comparable to that of human design. The overall performance surpasses existing automatic pipeline design methods, with an average throughput improvement of 19.0%. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
