Ginkgo OMP problem

[blegouix]

Yann and me observed numerical issues in Gysela with OMP, even with OMP_NUM_THREADS=1.

#319 imputes the issue to Ginkgo with OMP executor.

I reproduce the issue by replacing:

std::size_t constexpr ncells = 100;

with:

std::size_t constexpr ncells = 32;

in tests/splines/non_periodic_spline_builder.cpp

(Note: there is no problem with ncells = 10 too, which is the setup for the batched tests.)

Then the splines_tests_DEGREE_X_3_BSPLINES_TYPE_UNIFORM_EVALUATOR_POLYNOMIAL_ tests will fail with error sometimes > 1e-10.

Doing the same in the #319 branch leads to all tests passing, with error usually around 1e-16.

[blegouix]

Related branch:
https://github.com/CExA-project/ddc/tree/ginkgo-omp-debug

[tpadioleau]

Can you do a test without the preconditioner ? It was not clear it was helpful and it could help determining whether the problem comes from the solver or the preconditioner.

[blegouix]

You are right it is the preconditionner. Using 1u (diagonal Jacobi) is enough to remove the problem.

[tpadioleau]

see ginkgo-project/ginkgo#1563

[tpadioleau]

Could we consider switching to the GMRES solver in the meantime ?

[blegouix]

#377