The radixsort in the SDK isn't really going to help much.
Take a look at the bottom of the page here:
http://www.heterogeneouscompute.org/?page_id=7
There's an implementation you can download as well as a doc describing it in rough terms. It's not the fastest radix sort out there and it's very AMD specific (vector-width dependent, it won't work on nvidia architectures, but the same applies to nvidia's sorts).
Quicksort doesn't really make a lot of sense. Off the top of my head I think radix sort is the only sort that's O(N) in terms of the data set size, but it's also very specific to bit-wise sorting and hence is not a comparison sort.
