Reining in the Functional Verification of Complex Processor Designs with Automation, Prioritization, and Approximation

Our quest for faster and efficient computing devices has led us to processor designs with enormous complexity. As a result, functional verification, which is the process of ascertaining the correctness of a processor design, takes up a lion's share of the time and cost spent on making processors. Unfortunately, functional verification is only a best-effort process that cannot completely guarantee the correctness of a design, often resulting in defective products that may have devastating consequences.Functional verification, as practiced today, is unable to cope with the complexity of current and future processor designs. In this dissertation, we identify extensive automation as the essential step towards scalable functional verification of complex processor designs. Moreover, recognizing that a complete guarantee of design correctness is impossible, we argue for systematic prioritization and prudent approximation to realize fast and far-reaching functional verification solutions. We partition the functional verification effort into three major activities: planning and test generation, test execution and bug detection, and bug diagnosis. Employing a perspective we refer to as the automation, prioritization, and approximation (APA) approach, we develop solutions that tackle challenges across these three major activities. In pursuit of efficient planning and test generation for modern systems-on-chips, we develop an automated process for identifying high-priority design aspects for verification. In addition, we enable the creation of compact test programs, which, in our experiments, were up to 11 times smaller than what would otherwise be available at the beginning of the verification effort. To tackle challenges in test execution and bug detection, we develop a group of solutions that enable the deployment of automatic and robust mechanisms for catching design flaws during high-speed functional verification. By trading accuracy for speed, these solutions allow us to unleash functional verification platforms that are over three orders of magnitude faster than traditional platforms, unearthing design flaws that are otherwise impossible to reach. Finally, we address challenges in bug diagnosis through a solution that fully automates the process of pinpointing flawed design components after detecting an error. Our solution, which identifies flawed design units with over 70% accuracy, eliminates weeks of diagnosis effort for every detected error.PhDComputer Science & EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttps://deepblue.lib.umich.edu/bitstream/2027.42/137057/1/birukw_1.pd