https://hgpu.org/?p=6712
Belief Propagation by Message Passing in Junction Trees: Computing Each Message Faster Using GPU Parallelization