Non-linear Thinking with CUDA

Why is the naive solution n^3? Generating one partial sum costs n, and you need to do it n times, so n^2. Even if you need to store all the intermediate sums, you get them as you sum the entire partial sum. No memoization needed. What am I missing here?

I'm wondering if this problem can be approached as a standard convolution problem.