Recurrent CircuitSAT Sampling for Sequential Circuits

TL;DR

This paper presents a GPU-accelerated recurrent CircuitSAT sampling method for sequential circuits, enabling efficient generation of input stimuli with clock cycle constraints for digital hardware verification.

Contribution

It introduces a novel differentiable, GPU-accelerated sampling technique for sequential circuits, improving runtime efficiency and solution diversity over traditional SAT-based methods.

Findings

Achieves up to 105.1x runtime speedup over existing heuristics.

Effectively generates diverse valid input stimuli for benchmark circuits.

Demonstrates scalability and effectiveness on ISCAS-89 and ITC'99 benchmarks.

Abstract

In this work, we introduce a novel GPU-accelerated circuit satisfiability (CircuitSAT) sampling technique for sequential circuits. This work is motivated by the requirement in constrained random verification (CRV) to generate input stimuli to validate the functionality of digital hardware circuits. A major challenge in CRV is generating inputs for sequential circuits, along with the appropriate number of clock cycles required to meet design constraints. Traditional approaches often use Boolean satisfiability (SAT) samplers to generate inputs by unrolling state transitions over a fixed number of clock cycles. However, these methods do not guarantee that a solution exists for the given number of cycles. Consequently, producing input stimuli together with the required clock cycles is essential for thorough testing and verification. Our approach converts the logical constraints and temporal behavior of sequential circuits into a recurrent CircuitSAT problem, optimized via gradient descent to efficiently explore a diverse set of valid solutions, including their associated number of clock cycles. By operating directly on the circuit structure, our method reinterprets the sampling process as a supervised multi-output regression task. This differentiable framework enables independent element-wise operations on each tensor element, facilitating parallel execution during learning. As a result, we achieve GPU-accelerated sampling with substantial runtime improvements (up to 105.1x) over state-of-the-art heuristic samplers. We demonstrate the effectiveness of our method through extensive evaluations on circuit problems from the ISCAS-89 and ITC'99 benchmark suites.