Dropdims is surprisingly slow

[mcabbott]

I don’t have a solution, but notice this:

julia> r = rand(10,1,10);

julia> @btime dropdims($r, dims=2);
  529.307 ns (9 allocations: 224 bytes)

julia> nda = NamedDimsArray(r, (:x, :_, :y));

julia> @btime dropdims($nda, dims=:_);
  7.647 μs (18 allocations: 672 bytes)

julia> @btime (() -> dropdims($nda, dims=2))();
  7.533 μs (18 allocations: 672 bytes)

julia> @btime NamedDims.remaining_dimnames_after_dropping((:x, :_, :y), 2)
  1.419 ns (0 allocations: 0 bytes)
(:x, :y)

Not really related, but also notice that dropdims doesn’t match repeated names, perhaps it should?

julia> ndb = NamedDimsArray(ones(2,1,2,1), (:x, :_, :y, :_));

julia> dropdims(ndb, dims=:_)
2×2×1 NamedDimsArray{(:x, :y, :_),Float64,3,Array{Float64,3}}:
[:, :, 1] =
 1.0  1.0
 1.0  1.0

julia> dropdims(dropdims(ndb, dims=:_), dims=:_)
2×2 NamedDimsArray{(:x, :y),Float64,2,Array{Float64,2}}:
 1.0  1.0
 1.0  1.0

[nickrobinson251]

dropdims doesn’t match repeated names, perhaps it should?

I'm less worried about that one, since we treat operations on duplicate names as undefined behaviour.

[mcabbott]

Oh I don’t claim that’s a bug, sorry if that wasn’t clear, just something to think about. But at some point the behaviour is defined by how it behaves (or should be made an error if it’s never to be used). Anyway perhaps an orthogonal issue to the title.