Optimizing finite element matrix calculations using the general technique of element vectorization (Q1104044)

This paper deals with the vectorization of elementary stiffness matrix calculation phase in a finite element computation code. The author recalls the classical vectorization technique based on the vectorization over the Gauss integration points. He presents two limitations to this approach which are not general serious limitations, in my opinion. Then he presents his own approach. It is based on the subdivision of the triangulation in a member of a fixed size blocks of elements. The vectorization is achieved over the elements of each block. Better performance is then achieved when the blocks are bigger. On the other hand, since all the elements in a same block must have the same Jacobian matrix, this approach is, in my opinion, very specific to particular problems where the user can choose some particular triangulation. In the fourth section, the author gives three tests to illustrate the efficiency of his vectorized algorithm. He does not precise the triangulation used. He only says that it is composed of 256 elements. He tests his algorithm on a scalar and then on a vector processor. But he does not compare the performance of his algorithm with other algorithms as the classical one which he recalls in the beginning.