In this issue I document issues/misc thoughs while exploring the implementation of the high-level API for hybridized solvers. In particular, what I did is to combine the skeleton code proposed here with the developments so far in ExploringGridapHybridization.jl.
Issues/Changes I had to make in Gridap
Gridap.jl v0.17: MultiFieldCellField cannot be currently built out of fields which are defined on different triangulations, i.e., bulk versus skeleton. A check fails here. I commented out the check. See here
Generalization of CellQuadrature member variables in Gridap.jl see here
Generalization of CellPoint member variables in Gridap.jl see here
Discuss overloads of Reindex + VectorBlock see here
MultifieldFEComponentBasis assumes (it is hard-coded) that its cell array of arrays of Fields is created out of the result of applying BlockMap to an array of arrays of Fields. This prevents transformations required in Hybrid methods from [c][b] to [c][f][b]. Instead, following that workflow we first transform [c] into [c][f], and then [c][b][f], which is NOT equivalent to [c][b][f]. I converted the inner constructor to an outer constructor as a workaround. see here A more clean solution might be needed here.
Misc ideas
I think it would be useful in several contexts to have FaceToCellGlue parameterized by {Df,Dc}, where Df is the topological dimension of the face and Dc the topological dimension of the cell. Would it be also useful for other types of glues, such as, e.g., FaceToFaceGlue ? To-think.
In this issue I document issues/misc thoughs while exploring the implementation of the high-level API for hybridized solvers. In particular, what I did is to combine the skeleton code proposed here with the developments so far in
ExploringGridapHybridization.jl.Issues/Changes I had to make in Gridap
MultiFieldCellFieldcannot be currently built out of fields which are defined on different triangulations, i.e., bulk versus skeleton. A check fails here. I commented out the check. See hereCellQuadraturemember variables inGridap.jlsee hereCellPointmember variables inGridap.jlsee hereReindex+VectorBlocksee hereMultifieldFEComponentBasisassumes (it is hard-coded) that its cell array of arrays ofFieldsis created out of the result of applyingBlockMapto an array of arrays ofFields. This prevents transformations required in Hybrid methods from[c][b]to[c][f][b]. Instead, following that workflow we first transform[c]into[c][f], and then[c][b][f], which is NOT equivalent to[c][b][f]. I converted the inner constructor to an outer constructor as a workaround. see here A more clean solution might be needed here.Misc ideas
FaceToCellGlueparameterized by{Df,Dc}, whereDfis the topological dimension of the face andDcthe topological dimension of the cell. Would it be also useful for other types of glues, such as, e.g.,FaceToFaceGlue? To-think.