Reduced Quintic Implicit #77
Conversation
cwsmith
left a comment
cwsmith
left a comment
There was a problem hiding this comment.
Thank you @jvmespark .
In addition to the comments below, I have a a few requests:
- Please add a description in the PR. If you got the shape function/gradient/quadrature points from literature please reference it.
- The CMakeLists.txt file needs to be modified to add the new test.
- Please add a test to exercise
evaluate(...)(given user defined parametric coords in each element, return the field value at those locations). As with the integration test, I'd strongly prefer a test that we know the result for and can check against it (as opposed to a simpler 'it doesn't crash' test).
| const auto family = OMEGA_H_SIMPLEX; | ||
| auto len = 1.0; | ||
| return Omega_h::build_box(world, family, len, len, len, | ||
| 2, 2, (dim == 3 ? 2 : 0)); |
There was a problem hiding this comment.
I think the template dim can be removed and the third dimension hardcoded as zero.
| { | ||
| auto mesh2D = createMesh<2>(lib); | ||
| MeshField::OmegahMeshField<ExecutionSpace, 2, MeshField::KokkosController> omf2D(mesh2D); | ||
| runReducedQuinticImplicit<MeshField::KokkosController>(mesh2D, omf2D); |
There was a problem hiding this comment.
Is there an expected result from the integration? I'd much prefer a test that we know the result for.
There was a problem hiding this comment.
There is probably some test cases in M3DC1 that I can dig up and share with James.
src/MeshField_Integrate.hpp
Outdated
| 0.1167862757264 / 2.0), | ||
| IntegrationPoint(Vector3{0.249286745, 0.249286745, 0.501426509}, | ||
| 0.1167862757264 / 2.0), | ||
|
|
There was a problem hiding this comment.
I think some of the points here do not have enough precision and fail summing to 1.
There was a problem hiding this comment.
@jvmespark This looks like it is the standard Gauss-Legendre points. You can copy them from here with more precision: https://github.com/SCOREC/core/blob/6a972af6daceec626e589b4d4f63acc22e27317f/apf/apfIntegrate.cc#L216
Alternatively, it looks like we may be able to use this script in APF to calculate to machine precision: https://github.com/SCOREC/core/blob/6a972af6daceec626e589b4d4f63acc22e27317f/apf/apfPolyBasis1D.cc
| const auto [shp, map] = MeshField::Omegah::getReducedQuinticImplicitElement(mesh); | ||
| MeshField::FieldElement fes(mesh.nelems(), field, shp, map); | ||
|
|
||
| ReducedQuinticImplicitIntegrator<FieldElement> integrator(fes); |
There was a problem hiding this comment.
| ReducedQuinticImplicitIntegrator<FieldElement> integrator(fes); | |
| ReducedQuinticImplicitIntegrator integrator(fes); |
Either do not put in template or you have to include the templating of FieldElement. ie MeshField::FieldElement<templating>.
|
Can you add |
src/MeshField.hpp
Outdated
| }; | ||
|
|
||
| template <int ShapeOrder> auto getTriangleElement(Omega_h::Mesh &mesh) { | ||
| static_assert(ShapeOrder == 1 || ShapeOrder == 2); |
There was a problem hiding this comment.
| static_assert(ShapeOrder == 1 || ShapeOrder == 2); | |
| static_assert(ShapeOrder == 1 || ShapeOrder == 2 || ShapeOrder == 5); |
|
@cwsmith @Joshua-Kloepfer One thing that we noticed is that in core the integration only stores n-1 coordinates where n is the number of vertices. This is why |
|
@jacobmerson Currently in our code, we store all n coordinates so for Tets we decided to store all 4 parametric coordinates (https://github.com/SCOREC/meshFields/blob/jk/FixIntegration/src/MeshField_Integrate.hpp#L85) and in triangles we store 3 (https://github.com/SCOREC/meshFields/blob/jk/FixIntegration/src/MeshField_Integrate.hpp#L85). We felt that this was the right approach. Is there somewhere else where we are not storing all the entries? |
4 participants
No description provided.