(Face|Cell)(Scalar|Vector)Values -> (Face|Cell)Values

[fredrikekre]

Perhaps we can combine scalar and vector values into a single struct. The CellScalarValues and CellVectorValues structs are almost identical:

Ferrite.jl/src/FEValues/cell_values.jl

Lines 37 to 49 in 32d2a58

	struct CellScalarValues{dim,T<:Real,refshape<:AbstractRefShape} <: CellValues{dim,T,refshape}
	N::Matrix{T}
	dNdx::Matrix{Vec{dim,T}}
	dNdξ::Matrix{Vec{dim,T}}
	detJdV::Vector{T}
	M::Matrix{T}
	dMdξ::Matrix{Vec{dim,T}}
	qr::QuadratureRule{dim,refshape,T}
	# The following fields are deliberately abstract -- they are never used in
	# performance critical code, just stored here for convenience.
	func_interp::Interpolation{dim,refshape}
	geo_interp::Interpolation{dim,refshape}
	end

Ferrite.jl/src/FEValues/cell_values.jl

Lines 89 to 101 in 32d2a58

	struct CellVectorValues{dim,T<:Real,refshape<:AbstractRefShape,M} <: CellValues{dim,T,refshape}
	N::Matrix{Vec{dim,T}}
	dNdx::Matrix{Tensor{2,dim,T,M}}
	dNdξ::Matrix{Tensor{2,dim,T,M}}
	detJdV::Vector{T}
	M::Matrix{T}
	dMdξ::Matrix{Vec{dim,T}}
	qr::QuadratureRule{dim,refshape,T}
	# The following fields are deliberately abstract -- they are never used in
	# performance critical code, just stored here for convenience.
	func_interp::Interpolation{dim,refshape}
	geo_interp::Interpolation{dim,refshape}
	end

It should be possible to combine these into something like:

struct CellValues{N_t, dNdx_t, ...}
    N::Matrix{N_t}
    dNdx::Matrix{dNdx_t}
    dNdξ::Matrix{dNdξ_t}
    detJdV::Vector{T}
    M::Matrix{T}
    dMdξ::Matrix{dMdξ_t}
    qr::QR
    func_interp::IPF
    geo_interp::IPG
end

const CellScalarValues = CellValues{<: Number}
const CellVectorValues = CellValues{<: AbstractVector}
const CellMatrixValues = CellValues{<: AbstractMatrix}

We only use the distinction for e.g.

Ferrite.jl/src/FEValues/common_values.jl

Line 107 in 32d2a58

@propagate_inbounds shape_symmetric_gradient(cv::CellVectorValues, q_point::Int, base_func::Int) = symmetric(shape_gradient(cv, q_point, base_func))

which you can then instead dispatch on CellValues{<:AbstractVector}.

Question is if it is worth it.

[termi-official]

I kinda like the idea, as it allows to us address several open issues in a simpler way (e.g. complex-valued problems and general tensor-valued problems, ...) and we can reduce some code duplication. However, I am not super confident regarding the interaction with Tensors.jl here, as well as with constraints between the types which are necessary to avoid performance issues.