Boosting Verification Scalability via Structural Grouping and Semantic Partitioning of Properties

Abstract

From equivalence checking to functional verification to design-space exploration, industrial verification tasks entail checking a large number of properties on the same design. State-of-the-art tools typically solve all properties concurrently, or one-at-a-time. They do not optimally exploit subproblem sharing between properties, leaving an opportunity to save considerable verification resource via concurrent verification of properties with nearly identical cone of influence (COI). These high-affinity properties can be concurrently solved; the verification effort expended for one can be directly reused to accelerate the verification of the others, without hurting per-property verification resources through bloating COI size. We present a near-linear runtime algorithm for partitioning properties into provably high-affinity groups for concurrent solution. We also present an effective method to partition high-structural-affinity groups using semantic feedback, to yield an optimal multi-property localization abstraction solution. Experiments demonstrate substantial end-to-end verification speedups through these techniques, leveraging parallel solution of individual groups.

Incremental design-space model checking via reusable reachable state approximations

“…Preprocessing the models and properties, based on knowledge about the design space, before checking them with FuseIC3 may remove redundancies in the design space. Online property grouping algorithms [24] can by extended to dynamically reorder properties based on information reuse and semantic information learned during a model checking. Exploring synergies between offline and online property grouping algorithms is a promising research direction.…”

A Parallel Optimization Method for Robustness Verification of Deep Neural Networks

“…Borrowing from the partition mode, the verification problem can be automatically divided into multiple simpler sub-problems, and thus obtain high-quality tighter upper and lower bounds, which has been proven to be sound and complete [39]. It can divide the input space or neuron activation phases, with each worker v q responsible for the independent processing of a specific sub-problem [40] Ṽj := f , ψj , where q ∈ {1, . .…”