Main paper can be found here and their reference implementation is here.

efficiently bucketizing arrays based on keys
only small number of buckets are considered here (2 - 64)
for these cases, such an implementation will be much faster than full sort
Idea is to split the algo into 3 phases
- local - compute bucket histograms and their scans
- global - compute global location of each element
- local - compute local location of each element and perform scatter
They also perform local multisplit to reduce memory divergence!