How to efficiently access first element of all subarrays, i.e. [:,1] or explicitly [(:SWATI, :GWAT),:amt]

[fabern]

Hi, I was wondering if there ist an efficient way to access the first element of all the subarrays in a ComponentVector of ComponentVectors?

using BenchmarkTools
using ComponentArrays #, LabelledArrays, RecursiveArrayTools

SWATI=ComponentArray(amt = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0], conc = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.10])
GWAT =ComponentArray(amt = 1.0, conc = 3.0)
u0 = ComponentArray(SWATI = SWATI, GWAT = GWAT)

@btime sum(reduce(vcat,[u0.SWATI.amt, u0.GWAT.amt])) # 5.175 μs (54 allocations: 1.98 KiB)
@btime sum(vcat(u0.SWATI.amt, u0.GWAT.amt))          # 3.912 μs (48 allocations: 1.72 KiB)

Or would this require defining the ComponentVector with permuted dimensions? Further, I saw higher dimensional ComponentArrays mentioned in the README.md, but I have not seen any example how to generate them in the documentation. Could you provide an example of this, please?

[jonniedie]

There's a function called valkeys that can type-stabily create the keys of a ComponentArray wrapped in Vals for efficient iteration. Then you can just map over the keys and index into the ComponentArray. This will get you a Tuple of the values and you can sum over the inner levels then outer, if you want.

using BenchmarkTools, ComponentArrays

SWATI = ComponentArray(amt = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0], conc = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.10])
GWAT = ComponentArray(amt = 1.0, conc = 3.0)
u0 = ComponentArray(SWATI = SWATI, GWAT = GWAT)

julia> map(k -> view(u0, k).amt, valkeys(u0))
([1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0], 1.0)

julia> @btime sum(map(k -> sum(view($u0, k).amt), valkeys($u0)))
  4.333 ns (0 allocations: 0 bytes)
56.0

Note also that I'm using view(u0, k) here instead of u0[k] to avoid the allocation (getindex on ComponentArrays slices, unlike getproperty, which views).

[jonniedie]

As for the multidimensional ComponentArrays, that wouldn't work here because the number of elements of each of the subcomponents are different. If they weren't jagged like that, you could make a ComponentMatrix by just concatenating the ComponentVectors together

julia> SWATI = ComponentArray(amt = [1.0, 2.0], conc = [0.1, 0.2]);

julia> GWAT = ComponentArray(amt = [3.0, 4.0], conc = [0.3, 0.4]);

julia> u0 = [SWATI GWAT]
4×2 ComponentMatrix{Float64} with axes Axis(amt = 1:2, conc = 3:4) × FlatAxis()
 1.0  3.0
 2.0  4.0
 0.1  0.3
 0.2  0.4

julia> u0[:amt, :]
2×2 Matrix{Float64}:
 1.0  3.0
 2.0  4.0

julia> @btime sum(@view $u0[:amt, :]) # Also using @view here
  6.166 ns (0 allocations: 0 bytes)
10.0

[fabern]

Thanks for this helpful answer, Jonnie!

The first part solves my problem perfectly well.

For the second part regarding multidimensional ComponentMatrix: in your example the second axis is simply FlatAxis(). Is there a possibility to have both axes with named entires, e.g. allowing to access u0[:amt, :SWATI] ?