Distributing Only the Last Axis

The performance of some S3L functions can be enhanced by block-distributing only the last axis of the S3L array and making all other axes local. This rule applies to the FFT, sorting, and banded solver functions.

These functions are all optimized for operating on S3L arrays that are distributed in this manner. If an array that has a different type of distribution is being passed in as an argument, these functions automatically redistribute the array, perform the parallel computation and then restore it to its original form. Since this data redistribution introduces extra overhead, it is a good practice to ensure that S3L arrays passed to these functions follow this distribution plan.