Sorting

This chapter covers two types of sorting methods that rearrange keys (and their associated values) on GPUs into a desired order in parallel. The first type of sorting method is radix sort, which sorts keys by distributing them across buckets. We show that using the shared memory for output tiling and applying thread coarsening improves coalescing and reduces the overhead of the global exclusive scan. Although radix sort has the advantage of having a computational complexity that is lower than O(N*log(N)), it does not work for keys with complex ordering requirements. Therefore we also briefly look at the parallelization of comparison-based sorting that is applicable to general types of keys and comparison operations. A class of comparison-based sorting algorithms that is amenable to parallelization is merge sort. Merge sort can be parallelized by performing independent merge operations of different input segments in parallel as well as parallelizing within each merge operation. We also briefly discuss other popular parallel sorting algorithms.