Achieving Last-Mile Functional Coverage in Testing Chip Design Software Implementations

Ming Yan, Junjie Chen, Hangyu Mao, Jiajun Jiang, Jianye Hao, Xingjian Li, Zhao Tian, Zhichao Chen, Dong Li, Zhangkong Xian, Yanwei Guo, Wulong Liu, Bin Wang, Yuefeng Sun, Yongshun Cui

Abstract: Defective chips may cause huge losses (even disasters), and thus ensuring the reliability of chips is fundamentally important. To ensure the functional correctness of chips, adequate testing is essential on the chip design implementation (CDI), which is the software implementation of the chip under design in hardware description languages, before putting on fabrication. Over the years, while some techniques targeting CDI functional testing have been proposed, there are still a number of hard-to-cover functionality points due to huge input space and complex constraints among variables in a test input. We call the coverage of these points last-mile functional coverage.Here, we propose the first technique targeting the significant challenge of improving last-mile functional coverage in CDI functional testing, called LMT, which does not rely on domain knowledge and CDI internal information. LMT first identifies the relevant variables in test inputs to the coverage of last-mile functionality points inspired by the idea of feature selection in machine learning, so as to largely reduce the search space. It then incorporates Generative Adversarial Network (GAN) to learn to generate valid test inputs (that satisfy complex constraints among variables) with a larger possibility. We conducted a practical study on two industrial CDIs in Huawei to evaluate LMT. The results show that LMT achieves at least 49.27% and 75.09% higher last-mile functional coverage than the state-of-the-art CDI test input generation techniques under the same number of test inputs, and saves at least 94.24% and 84.45% testing time to achieve the same functional coverage.