Consider Setting Default OpenMP schedule(static) for Execution Policies

[rchen20]

Is your feature request related to a problem? Please describe.

On TOSS4 Intel systems, OpenMP sum reductions in the Launch tests fail with inaccurate answers:

325 - test-launch-basic-param-expt-ReduceSum-OpenMP.exe (Failed)
326 - test-launch-basic-param-expt-ReduceMin-OpenMP.exe (Failed)

Changing the OpenMP schedule to static or dynamic resolves the problem. For example, changing this line

RAJA/include/RAJA/policy/openmp/launch.hpp

Line 182 in 1ddae3d

#pragma omp for

to:

#pragma omp for schedule(static)

gives the correct answers in the tests. Note that setting the schedule to guided, runtime, or auto begets failures.

Describe the solution you'd like

Currently, the omp_for_exec policy either doesn't specify a schedule, or uses the auto schedule (in the forall case). We should consider setting this policy to automatically use schedule(static).

Additional context

We are also in contact with Greg Lee from LC about the failing tests. He has submitted issue 06375267 to Intel.

[rchen20]

Greg Lee suggested compiling with -fp-model-precise which also solves the problem.

[rhornung67]

I don't think we should change our internal implementation to something that conflicts with the OpenMP standard. Rather, I think it is best to make the need to be explicit about this clear in our User Docs about execution policies. I will take a swing at this and put up a PR.

[rchen20]

This is being handled here #1755.