This is the implementation of a k-way multilevel partitioner, an extension of the sorting algorithm presented in the eponymous paper.
We present a partitioning algorithm that works in-place, is cache-efficient, avoids branch-mispredictions, and performs work O(n log k) for arbitrary inputs. The main algorithmic contribution is a coarse-grained block-based permutation which makes distribution-based algorithms in-place.
#include "ips2pa.hpp"
// sort sequentially
ips2pa::partition(It begin, It end, It s_begin, It s_end, size_t* b_begin, bool use_equal_buckets[, Comp comp, std::ptrdiff_t seed])The parameters begin and end determine the input data, the parameters s_begin and s_end describe the splitters, and the parameter b_begin points to the output array which stores the bucket sizes. If the parameter use_equal_buckets is set to true, the partitioner groups elements together which are equal to splitters.
Make sure to compile with C++11 support. Currently, the code does not compile on Windows.