Approaches for parallelizing reductions on modern GPUs