Conversation
…mfem 4.8. This primary change is to rename the tps RiemannSolver class to RiemannSolverTPS. This is done b/c mfem has introduced a RiemannSolver into the mfem namespace, and tps is very promiscuous about using namespace mfem, leading to a clash. We could (probably should) fix that, but renaming the tps RiemannSolver is quicker.
In mfem v4.7 the ElementTransformation::mesh data member was changed to const. We use this in the GradientVectorGridFunctionCoefficient::Eval function, which reproduces some code from mfem::RefinedToCoarse (see mfem/fem/coefficient.cpp). So, for mfem 4.7 and later we require const here. But, prior to mfem v4.7, we cannot use const here b/c the subsequent function call to mesh::GetRefinementTransforms() violates const-correctness. So, the solution is the #if in this commit.
mpi4py = 3.1.4, to avoid issues with underlying mpi? mfem = 4.6, to avoid tps test failure with 4.8 that is under investigation
|
Some notes about failing tests...
The diffs are smallish (e.g., max absolute diff in density of ~5e-9, corresponding to a relative diff of roughly ~1e-7) but not small enough to automatically attribute to double precision arithmetic without further investigation. I am currently trying to further pin down what mfem change between v4.6 and v4.6.2-rc0 leads to this behavior. |
It appears that the first point at which the tests fail is when mfem PR 3899 is merged. This PR appears to include a slight modification in the detection of boundary points. This could be the cause, but still digging into it. |
This is the cause. Local testing shows that prior to mfem PR 3899, the the tests pass as is and that very few "boundary" points are detected in the gslib-based interpolation. I think this is because gslib is only detecting points on actual boundaries rather than interelement boundaries, but it is hard to parse the gslib code so that is a tentative conclusion. Either way, after mfem PR 3899 is merged, many more boundary points are detected in the interpolation, and To address this, since I'm not ready to move to mfem 4.8 all test environements (since most folks are still using mfem 4.5.2), logic has been added to these tests to detect the mfem version and use an appropriate reference solution. This should allow the regression tests to pass regardless of the mfem version. |
2 participants
This PR implements minor changes required to upgrade to mfem 4.8.