ecms_neu_mini.png

Digital Library

of the European Council for Modelling and Simulation

 

Title:

Vectorized Implementation Of The FEM Numerical Integration Algorithm On A Modern CPU

Authors:

Filip Kruzel

Published in:

 

 

(2019). ECMS 2019 Proceedings Edited by: Mauro Iacono, Francesco Palmieri, Marco Gribaudo, Massimo Ficco, European Council for Modeling and Simulation.

 

DOI: http://doi.org/10.7148/2019

 

ISSN: 2522-2422 (ONLINE)

ISSN: 2522-2414 (PRINT)

ISSN: 2522-2430 (CD-ROM)

 

33rd International ECMS Conference on Modelling and Simulation, Caserta, Italy, June 11th – June 14th, 2019

 

 

Citation format:

Filip Kruzel (2019). Vectorized Implementation Of The FEM Numerical Integration Algorithm On A Modern CPU, ECMS 2019 Proceedings Edited by: Mauro Iacono, Francesco Palmieri, Marco Gribaudo, Massimo Ficco European Council for Modeling and Simulation. doi: 10.7148/2019-0414

DOI:

https://doi.org/10.7148/2019-0414

Abstract:

The main aim of this study is to answer the question: how to e˙ectively implement the creation of the finite element sti˙ness matrix in parallel simulations of Fi-nite Element Method using the full advantages of mo-dern multiprocessors such as parallelization combined with vectorization. In this work, an eÿcient method for implementation of a Finite Element Method numerical integration algorithm on a modern Intel Haswell CPU architecture was developed. This algorithm was cho-sen, due to its non-trivial structure and the fact that its optimization is often omitted in research in favour of accelerating the other phases of FEM. Tests included two types of tasks to solve, with the use of two ty-pes of approximation and two types of finite elements. During this study, several methods for the implementa-tion of the chosen algorithm was investigated, including Intel Cilk Plus, Intel Intrinsics and other computing techniques. Results were compared with an older San-dy Bridge architecture, showing a significant impact of vectorization and large cache on the performance of the modern CPUs. Our research gives suggestions for cho-osing the optimal design of algorithms and e˙ectively using all of the features of the modern CPUs.

Full text: