high performance computing on graphics processing units: hgpu.org

We present an investigation of the use of GPGPU techniques to parallelize the execution of a satisfiability solver, based on the traditional DPLL procedure – which, in spite of its simplicity, still represents the core of the most competitive solvers. The investigation tackles some interesting problems, including the use of a predominantly data-parallel architecture, like NVIDIA’s CUDA platform, for the execution of relatively "heavy" threads, associated to traditionally sequential computations (e.g., unit propagation), non-deterministic computations (e.g., variable splitting), and meta-heuristics to guide search. Experimentation confirms the potential for significant speedups from the use of GPGPUs, even with relatively simple modifications to the structure of the DPLL procedures – which should facilitate the porting of such ideas to other DPLL-based solvers.