AMG Direct solver on GPU for windows

[eclipseyin147]

It seems that windows openmp is not supported across the project, then direct solver is also not supported for AMG direct solver at the coarsest level.
So how can I run a direct solver on windows GPU ?

Thanks a lot !

[yhmtsai]

@eclipseyin147 Which GPU do you have on Windows? If you can make corresponding toolkit works with your GPU, you might compile Ginkgo's GPU backend successfully. Unfortunately, we do not have the testing server with these GPU for windows, so I can not give you any guarantee on that.

[eclipseyin147]

I have NVIDIA 3070 here. I compiled it with vcpkg on windows. Actually it can run AMG preconditioner on my GPU only when I set the coarsest solver to CG rather than the direct solver . Like this:

auto smoother =
                        gko::share(
                            gko::preconditioner::Jacobi<SolverReal, int32_t>::build()
                            .on(exec));
                auto coarsest_solver =
                        gko::solver::Cg<SolverReal>::build()
                        .with_criteria(gko::stop::Iteration::build()
                            .with_max_iters(100u)
                            .on(exec))
                        .on(exec);
                auto amg = gko::solver::Multigrid::build()
                        .with_mg_level(
                            gko::multigrid::Pgm<SolverReal, int32_t>::build()
                            .with_deterministic(false)
                            .with_max_iterations(10u)
                            .with_max_unassigned_ratio(0.1))
                        .with_max_levels(20u)
                        .with_min_coarse_rows(500u)
                        .with_pre_smoother(smoother)
                        .with_post_smoother(smoother)
                        .with_coarsest_solver(std::move(coarsest_solver))
                        .with_criteria(
                            gko::stop::Iteration::build().with_max_iters(1u).on(
                                exec))
                        .on(exec)
                        ->generate(matrix);
                return gko::solver::Fcg<SolverReal>::build()
                        .with_generated_preconditioner(gko::share(std::move(amg)))
                        .with_criteria(std::move(openFoamCriterion), std::move(maxIter))
                        .on(exec)
                        ->generate(matrix);

[yhmtsai]

you can remove .with_coarsest_solver(std::move(coarsest_solver)), it will use direct solver by default on non-distributed matrix