Chapter 8.5 : How to optimize more
The main issue in a reduction is called back dependency.
In other words, you have to wait until the end of the computing of the current element to compute the next one.
Figure 20 : Illustration of the computing elapsed time between the computation and the moment the result can be used. And, how to take account this particularity to optimise the computation time of the reduction.
To optimise the reduction computation we are going to use several accumulators to do the reduction.