Improve velocity with use of GPU instead of cpu with tools like OGL

[stefano2734]

Description of feature

GPU with using OpenCL, HIP and CUDA can accelerate about 4 to 100 times against cpu solution.

Potential solution

OGL Openfoam Ginkgo Layer with CUDA can do this with Nvidia Hardware.

See https://github.com/hpsim/OGL

minimum openfoam version is here 2106. so an update is here necessary in openfoam base.

and a solution with OpenCL is also possible with openfoam 1912 with solver petsc4foam

https://www.esi-group.com/sites/default/files/resource/other/1806/8th_OpenFOAM_Conference_Cineca_Spisso_2_0.pdf

https://wiki.openfoam.com/images/c/cd/OpenFOAM_2020_KAUST_Zampini.pdf

third is OpenMP 5.0 and higher (actual 5.2, 6.0 wip) with possible offloading to GPU with newer adapted openfoam
https://arxiv.org/pdf/2205.05681.pdf
https://www.openmp.org/wp-content/uploads/OpenMP_ECP_Sollve_May2023_AMDtalk.pdf

https://www.openmp.org/resources/refguides/

So two ways: external layer or internal in openfoam with newer Version of openfoam.

actual installation nextfoam with actual openfoam 23.06 and petsc4foam
https://blog.nextfoam.co.kr/2024/01/10/gpu-accelerated-openfoam-with-petsc4foam/

[friedenhe]

@stefano2734 Thanks for the suggestion. We do want to upgrade the OpenFOAM version to the latest. The main limitation is that we need to differentiate the newer OF version using CoDiPack and MeDiPack. This process is very time-consuming (on the order of months), and we do not have the manpower to do that at this moment...

[greole]

Hi @stefano2734 @friedenhe I just stumbled over this repo and issue. If you need any help with getting OGL for this repo to work, let me know. Also we are working with a somewhat larger team and scope on a rewrite of core routines for finite volume cfd NeoN. A hybrid solution using NeoN and OpenFOAM is shown for icoFoam neoIcoFoam. It is WIP and we're working on mpi support and simplifying things at the moment, but it already features full GPU support.