Regression in BroadcastQuasiArray about continuumarrays.jl HOT 5 CLOSED

Hmm, the IndexLinear stuff is a bit ill-defined for QuasiArrays.jl. So I'm wondering if this should just be removed as a concept?

Probably for now we just need to add an error_if_canonical_getindex overload

from continuumarrays.jl.

Actually, I'd be happy if I could just evaluate any function, however complicated, as long as it's defined on the Inclusion and returns a scalar. It feels I too often have to second-guess what QuasiBroadcastArray can handle.

For full context: to implement absorbing boundary conditions, I use a imaginary or complex potential at the far end of the box. It's a simple function that smoothly goes from zero to large imaginary part.

Here is a MWE (depends on Elliptic.jl):

using CompactBases
using IntervalSets
import Elliptic.Jacobi: cn

struct ComplexAbsorbingPotential{T,I}
    Emin::T
    kmin::T
    δ::T
    Δr::T
    interval::I
end

capx(cap::ComplexAbsorbingPotential) =
    r -> 2cap.δ*cap.kmin*r

capy(cap::ComplexAbsorbingPotential) =
    x -> -im*cap.Emin*√(cn(x/√2,1/2)^(-4) - 1)

R = StaggeredFiniteDifferences(100, 0.5)
cap = ComplexAbsorbingPotential(1.0, √2, 0.1, 9.27037338650686,
                                40.97962661349314..50.25)
r = axes(R,1)
i = cap.interval

a,b = i.left,i.right

# Translate cap to interval
x = capx(cap).(r .- a)
v = (a .< r .≤ b) .* capy(cap).(x)

R'*QuasiDiagonal(v)*R

which fails with an error similar to the above.

EDIT: Yet another example:

using IntervalSets
using QuasiArrays
using ContinuumArrays

x = Inclusion(0.0..20.0)
exp.(x)[2.0] # Works
exp.(-x)[2.0] # Crashes
exp.(x.^2)[2.0] # Crashes
exp.(-x.^2/(20/2))[2.0] # Crashes

from continuumarrays.jl.

Super, thanks! I was just about to push my own fix, but this looks better.

from continuumarrays.jl.

Can you see if that branch fixes all the issues?

from continuumarrays.jl.

Indeed, it works perfectly!

from continuumarrays.jl.