Performance of OpenMP loop transformations for the acoustic wave stencil on GPUs

Jaime Freire de Souza (1, 2) , Leticia Suellen Farias Machado (1) , Edson Satoshi Gomi (2) , Claude Tadonki (3, 4) , Simon Mcintosh-Smith (5) , Hermes Senger (1, 2)

1 UFSCar - Universidade Federal de São Carlos [São Carlos]
2 Universidade Federal de São Paulo
3 Mines Paris - PSL (École nationale supérieure des mines de Paris)
4 CRI - Centre de Recherche en Informatique
5 University of Bristol [Bristol]

Main Findings : • As a general remark, both loop transformations, unroll and tiling can yield significant improvements to the performance of the kernel evaluated on all GPUs evaluated. • Performance gains ranged from 1.13x to 2.93x. In most scenarios, the best performance was achieved by combining unroll and tiling. • The performance of tiling is highly sensitive to the choice of block size.

Domaines: Informatique [cs]; Performance et fiabilité [cs.PF]; Calcul parallèle, distribué et partagé [cs.DC]

Jaime Freire de Souza, Leticia Suellen Farias Machado, Edson Satoshi Gomi, Claude Tadonki, Simon Mcintosh-Smith, et al.. Performance of OpenMP loop transformations for the acoustic wave stencil on GPUs. SC22 The International Conference for High Performance Computing, Networking, Storage, and Analysis, Nov 2022, Dallas, United States. . ⟨hal-03888100⟩