Related write-ups below along with comparison of performance of different approaches. BMI2 instruction set based method seems to be the fastest of the lot.
I used it to accelerate nearest neighbor detection for collision processing in particle-laden flow for modeling complicated domains in 3d (for biological fluids simulation). I was using it as a locality sensitive hashing to put particles near each other in the same bucket in an hash map. I came across the ideas of BIGMIN (big minimum) and LITMAX (little maximum) for range search in a morton encoded data that I found to be cool.