Approximation-aware rewriting of AIGs for error tolerant applications
Approximation-aware rewriting of AIGs for error tolerant applications
复制标题
DOI：
10.1145/2966986.2967003
复制
发表时间：
2016-11
期刊：
影响因子：
--
通讯作者：
Arun Chandrasekharan;Mathias Soeken;Daniel Große;R. Drechsler
中科院分区：
文献类型：
--
作者：
Arun Chandrasekharan;Mathias Soeken;Daniel Große;R. Drechsler
与传统电路相比，近似电路以计算准确性为代价提供了出色的性能（速度和面积）。根据几个误差指标，例如最差案例误差，射线误差或误差率，评估了近似电路中结果的准确性。多个应用程序的误差指标要求有所不同，即，必须一次或组合满足所有错误标准。然而，所有应用程序都受益于改善延迟和面积。具有正式保证错误指标的自动合成方法对产生符合这些标准的电路非常有帮助。此外，这些指标中的每一个都是独立的数量（一个指标的价值与另一个度量无关），并且自动合成可以发现一个或多个放松度量的机会，对另一个或多个放松的指标进行了严格的要求，从而可以提高性能。 。在本文中，我们使用基于逆变器图（AIG）重写的自动合成方法，不仅可以改善性能，还可以保证引入近似错误的边界。我们在广泛的设计和标准基准电路上评估了我们的合成方法，以显示有用性和适用性。特别是，我们表明我们的合成结果甚至可以与手工制作的Adhoc近似电路（例如近似添加剂）在图像压缩的案例研究中获得的优化相媲美。
Approximation circuits offer superior performance (speed and area) compared to traditional circuits at the cost of computational accuracy. The accuracy of the results in approximation circuits is evaluated based on several error metrics such as worst-case error, bit-flip error, or error-rate. Several applications have varied requirements in error metrics, i.e., all the error criteria have to be met together at a time, or in combinations. Nevertheless, all applications benefit from improved delay and area. An automated synthesis approach with formal guarantees on error metrics is very helpful in generating circuits that meet these criteria. Furthermore, each of these metrics are independent quantities (value of one metric does not correlate with the other), and automated synthesis can discover opportunities to trade off one or more of the relaxed metrics with a strict requirement on the other, resulting in better performance. In this paper, we present an automatic synthesis approach using And-Inverter Graphs (AIGs) based rewriting that not only improves the performance but also guarantees the bounds of approximation errors introduced. Our synthesis approach is evaluated on a wide range of designs and standard benchmark circuits to show the usefulness and applicability. In particular, we show that our synthesis results are even comparable with the optimization achieved with hand crafted adhoc approximation circuits such as approximation adders in a case study on image compression.