Automatic Generation of System Level Assertions from Transaction Level Models
Automatic Generation of System Level Assertions from Transaction Level Models
复制标题
DOI：
10.1007/s10836-013-5403-y
复制
发表时间：
2013-09
期刊：
影响因子：
--
通讯作者：
Lingyi Liu;Shobha Vasudevan
中科院分区：
文献类型：
--
作者：
Lingyi Liu;Shobha Vasudevan
我们从事务级模型 (TLM) 模拟跟踪中自动生成断言。生成的断言以具有定量时间约束的线性时间逻辑的形式表达设计规范[4]。我们首先生成断言，而不考虑定量的时间限制。它们以模拟痕迹中的频繁模式的形式被挖掘。我们使用情节挖掘来挖掘模拟痕迹，以识别包含函数调用和事件的频繁情节。然后，我们用实时参数来注释剧集，以表达剧集中函数调用或事件之间的定量时间约束。在挖掘此类 TLM 断言时，我们采用符号执行来概括跟踪中函数调用的参数和返回值，以帮助挖掘引擎生成高质量的断言。我们构建了一个现实的基于 AXI 的互连网络平台，并在其上展示了实验结果。我们证明，我们的技术可以在基于 AXI 的平台以及基于 AMBA 的事务级 DMA 控制器上有效地生成高质量的性能和功能断言。我们证明，与之前使用顺序模式挖掘的努力相比，情节挖掘更具可扩展性，并且能够生成更紧凑的高质量 TLM 断言集。使用episode挖掘生成的断言数量最多可以减少228倍，并且每个断言中两个事件/函数调用之间的时间间隔小于50个时间单位。
We automatically generate assertions from Transaction Level Model (TLM) simulation traces. The generated assertions express design specifications in the form of linear temporal logic with quantitative temporal constraints [4]. We first generate the assertions without regard to the quantitative time constraints. They are mined in the form of frequent patterns in the simulation traces. We mine simulation traces usingepisode miningto identify frequent episodes comprising function calls and events. We then annotate the episodes with real time parameters to express quantitative time constraints among the function calls or events in the episode. When mining such TLM assertions, we employ symbolic execution to generalize the parameters and return values of function calls in the traces to help the mining engine generate high quality assertions. We have constructed a realistic AXI-based interconnection network platform that we demonstrate experimental results on. We show that our technique efficiently generates high quality performance and functional assertions on the AXI-based platform as well as a transaction level AMBA-based DMA controller. We demonstrate that episode mining is more scalable and able to generate a more compact set of high quality TLM assertions than previous efforts using sequential pattern mining. The number of generated assertions using episode mining can be reduced by up to 228 times, and the time interval between two events/function calls in each assertion is smaller than 50 time units.