Fix wrong usage of shape_value_type

[KnutAM]

Want to see test coverage first, and then should add test that catches this (unless I thought something wrong here - could you check @lijas?)

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 93.75%. Comparing base (6eead25) to head (5f0b0f8).

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1160   +/-   ##
=======================================
  Coverage   93.74%   93.75%           
=======================================
  Files          39       39           
  Lines        6238     6240    +2     
=======================================
+ Hits         5848     5850    +2     
  Misses        390      390           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[lijas]

The function shape_value_type and your newly added _function_value_type does basically the same thing, so i dont understand what this fixes?

[KnutAM]

The way I understand is that

• shape_value_type(ip, T): "Return the type of shape_value(ip::Interpolation, ξ::Vec, ib::Int)." (from docstring, should be reference_shape_value and ξ::Vec{rdim, T} AFAIU)
• _function_value_type: Returns the type of sum(i -> reference_shape_value(ip, ξ, i) * u[i], 1:getnbasefunctions(ip))

So the difference would occur if the eltype of u would be a tensor, because shape_value_type(ip::ScalarInterpolation, T::Type{<:AbstractTensor}) is not defined, while function_value_type(ip::ScalarInterpolation, ::AbstractVector{<:AbstractTensor}) is.

I was surprised so perhaps we are never supposed reach this case in the code here (discovered it when working with #1161 where such a function became useful for the same purpose as used here).

[lijas]

Ah I see, _evaluate_at_grid_nodes can evaluate with tensors also (not only the solution vector). Yes, then this makes sense.

[fredrikekre]

Can function_value_init be used?

[KnutAM]

@fredrikekre's review comments