Question

The application is to intersect two sorted list of integers (set intersection), say list1 and list2.

Each element of list1 will be assigned a GPU thread, and do binary search to check whether it appears in the list2. It is easy to see that there will be huge amount thread divergences in this application. I wonder if there is any good approach to reduce thread divergences. I am using CUDA to implement this application.

I know there is an approach called P-ary search, but my task is to reduce the thread divergence of binary search. Also I know there is library called thrust, but it seems there is no attempt on reducing the divergences.

Answer 1

If both lists are sorted, binary search is not the best algorithm you can do. Binary search will give O(n lg n), but just doing a merge-like algorithm, only taking intersections, is O(n).

This is a silly algorithm to use a GPU for. The only case I see is that you’ve just generated the data in the GPU. In which case, you want to break the problem up into a bunch of smaller intersections and assign a thread to each.

To do that, pick k equally-spaced elements of list1 and find them in list2 using binary search. Similarly, pick k equally-spaced elements of list2 and find them in list1. You now have 2k ranges in each list, where each range has at most N/k elements. Now intersect those ranges in parallel. (Set k to be half the number of threads you want.)