Hands-On GPU Programming with Python and CUDA

Work-efficient parallel prefix (down-sweep phase)

Now let's continue with the down-sweep, which will operate on the output of the up-sweep:

input: x0, ..., xn-1initialize:    for i = 0 to n - 2:        yi := xi    yn-1 := 0begin:for k = log2(n) - 1 to 0:    parfor j = 0 to n - 1:         if j is divisible by 2k+1:            temp := yj+2k-1            yj+2k-1 := yj+2k+1-1            yj+2k+1-1 := yj+2k+1-1  temp        else:            continueendoutput: y0 , y1 , ..., yn-1