Por: Pedro Areias

Universidade de Évora

Data: 5 de Junho 2013

Hora: 14:30
Local: Anfiteatro 1 - CLAV

Abstract:

Interconnected multiple point constraints can be topologically ordered and enforced by dense matrix multiplication as recently shown by Areias et al. where a Gaussian sparse solver was used. For equality constraints, both the solution and all intermediate reactions are obtained directly from the solution of a reduced linear system. However, standard sparse solvers require the additional step of symbolic assembling and typically do not make use of dense linear algebra kernels: indirect addressing is required during decomposition, which limits the use of dense kernels and effectiveness of shared-memory parallelism. In this work a new approach is presented for the frontal solution method within the multiple point constraint framework. An a-priori pivot and front sequence is established. Some of the constraint enforcement calculations are performed during the (local) assembling stage. This has been found to speed up forward elimination and back-substitution and directly conforms to the concepts set up by our previous work. Element ordering makes use of a variant of Sloan's algorithm where the first stage uses exact level sets. When compared with both standard sparse solvers and classical frontal implementations, memory requirements and code size are significantly reduced. Examples of clique-based problems are shown with large systems being solved in core.