CUDA - why is warp based parallel reduction slower?

Question

I had the idea about a warp based parallel reduction since all threads of a warp are in sync by definition.

So the idea was that the input data can be reduced by fact

Answer 1

You should also check the Examples in the SDK. I remember one very nice example with implementations of several ways of reductions. At least one of those also uses warp based reduction.

(I can't look up the name right now, because I have it only installed on my other machine)