Accelerating Exact and Approximate Inference for (Distributed) Discrete Optimization with GPUs