Cache‐efficient renumbering for vectorization

Abstract

A renumbering strategy for field solvers based on unstructured grids that avoids memory contention and minimizes cache‐misses is described. Compared with usual colouring techniques, the new renumbering strategy reduces the spread in point‐data access for edge‐based solvers by more than an order of magnitude. The technique is particularly suited for multicore, cache‐based machines that allow for vectorization or pipelining.

Full Document