A fast implementation of a spectral finite elements method on CPU and GPU applied to ultrasound propagation

Carlos Carrascal-Manzanares (1, 2) , Alexandre Imperiale (2) , Gilles Rougeron (2) , Vincent Bergeaud (2) , Lionel Lacassagne (1)

1 ALSOC - Architecture et Logiciels pour Systèmes Embarqués sur Puce
2 DISC (CEA, LIST) - Département Imagerie et Simulation pour le Contrôle (CEA, LIST)

In this paper we present an optimization of a spectral finite element method implementation. The improvements consisted in the modification of the memory layout of the main algorithmic kernels and in the augmentation of the arithmetic intensity via loop transformations. The code has been deployed on multi-core SIMD machines and GPU. Compared to our starting point, i.e. the original scalar sequential code, we achieved a speed up of ×228 on CPU.We present comparisons with the SPECFEM2D code that prove the good performances of our implementation on similar cases. On GPU, a hybrid solution is investigated.

Mots clés: parallelization; SIMD; GPU; vectorization; Spectral Finite Elements

Domaines: Calcul parallèle, distribué et partagé [cs.DC]; Performance et fiabilité [cs.PF]; Architectures Matérielles [cs.AR]; Mathématique discrète [cs.DM]; Algorithme et structure de données [cs.DS]; Modélisation et simulation; Analyse numérique [cs.NA]

Carlos Carrascal-Manzanares, Alexandre Imperiale, Gilles Rougeron, Vincent Bergeaud, Lionel Lacassagne. A fast implementation of a spectral finite elements method on CPU and GPU applied to ultrasound propagation. International Conference on Parallel Computing, ParCo Conferences, Sep 2017, Bologna, Italy. pp.339 - 348, ⟨10.3233/978-1-61499-843-3-339⟩. ⟨hal-01761605⟩