I have a recursive O(n log(n)) (I think) sorting algorithm that should work decently if parallelized. Here is the code using C++ code.

// Groups the array into two parts: low and high. for(int i = 0; i < LENGTH/2; i++) { auto &objectA = vec[i]; auto &objectB = vec[(LENGTH+1)/2+i]; if(objectB < objectA) std::swap(objectA, objectB); } // Exposes the highest half of low and lowest half of high. // The next two statements can be done concurrently std::partial_sort (vec.begin() + LENGTH/2, vec.end() - LENGTH/4, vec.end()); std::partial_sort (vec.rbegin() + LENGTH/2, vec.rend() - LENGTH/4, vec.rend(), [](int a, int b){ return b<a; }); /* Simplified from the these two sort calls. * Sorts the exposed halves. Makes low and high only have the lowest and highest entries, respectively. * std::sort (vec.begin() + LENGTH/4, vec.end() - LENGTH/4); * Sorts the low half of the array. * std::sort (vec.begin(), vec.begin() + LENGTH/2); */ std::sort (vec.begin(), vec.end() - LENGTH/4); // Sort the high half of the array. std::sort (vec.end() - LENGTH/2, vec.end());

There are four recursive calls so I'm getting a headache by looking at it. I don't have experience simplifying very complex algorithms like. After dealing with the recursion, my main concern is what to do when the arrays get small. Sorting those small arrays could take a serious chunk out of performance.

Any ideas for the algorithm? Any ideas on how I should make this more comprehensible for myself?

EDIT: When replacing the std::sorts with my own implementation, I realized I used a sum instead of a product to calculate the big O.

Message was edited by: Kyle Siefring

Can you explain, the algorithm a bit more? And indicate what part of it is parallelizable as per you.

I don't think you need to consider the situations when input arrays are small, as most likely you will not use OpenCL for that. Have you checked Radix Sort & Bitonic Sort ?