I am in need to optimize the performance of the 3D solver that integrates a system of equations for seismic wave propagation. Unsurprisingly, the function that implements the finite-difference stencil computations, is the bottleneck of the computations.
I read this article, where in Section 3.2 the authors say that the register blocking allowed them to obtain substantial speedup of the 3D stencil computations.
I do not understand how to implement this technique. Is it possible to find somewhere the description of register blocking in clear form?