https://hgpu.org/?p=6113
Optimizing and Auto-tuning Belief Propagation on the GPU