Hardware Acceleration for Constraint Solving for Random Simulation

Constrained random simulation has been widely adopted in contemporary hardware verification flows. In this methodology, a set of user-specified declarative constraints describe valid input stimuli for the design under test (DUT). A constraint solver produces the simulation input vectors; their generation is interleaved with the actual simulation of the design for these vectors. Besides its distribution, the solver´s performance is one of the most critical characteristics that determines the overall verification efficiency. There are no general approaches to hardware acceleration for solving declarative constraints. Current setups for hardware acceleration-based verification combine a software constraint solver running on a general-purpose processor with the hardware-accelerated DUT. This approach suffers from a major efficiency bottleneck caused by the significant performance mismatch between the solver executed in software and the DUT running on an accelerator. In this paper, we present a hardware constraint solver that uses a set of parallel solving units executing Markov chain Monte Carlo sampling. We propose to combine this solver and the DUT on the same device and run both entities hardware-accelerated in order to eliminate the performance mismatch. We discuss the details of the solver architecture and its implementation and report comprehensive results on performance and distribution characteristics as well as experience obtained from our case study where we used our solver to verify a real-world hardware design.