I am writing a simulation that outputs a multidimensional array A that changes over time, and whose ndims depends on the input dimension (1 or 2 for now). I would like to save A over time in a larger array A_t in my for-loop, for which I used CartesianIndices. The following is a MWE: In 1D: A_t = …

That’s a really ugly hack. Why not do A_t[fill(:, 2)..., 1] = A instead?

Use $ to flag the non-global “arguments” to your benchmarked expression in order to get a true sense of the performance here. julia> @btime A_t[CartesianIndices(A),1] = A; 157.917 ns (1 allocation: 16 bytes) julia> @btime A_t[CartesianIndices(A),1] .= A; 1.449 μs (8 allocations: 176 bytes) ju…

This confused me a bit at first, too, since: julia> A_t[CartesianIndices(A), 1] 10×10 Array{Float64,2}: 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 …

[image] simeonschaub: So getindex actually materializes into a matrix, but when using setindex! with CartesianIndices , it behaves like assigning to a one-dimensional array. Note that when you use square-bracketed indexing as the left hand side of a = assignment, it’s not doing getindex (and …

[image] simeonschaub: A[:] basically flattens A into a one-dimensional Vector , so the dimensions of the left and right hand side now match. This is still significantly faster than broadcasting on my machine. Thank you! This is indeed faster than broadcasting for me too. And in 1D, A_t = zer…

[image] mbauman: Note that when you use square-bracketed indexing as the left hand side of a = assignment, it’s not doing getindex (and thus isn’t materializing the array). I understand that, but intuitively, one would expect, that if getindex returns a matrix, setindex with the same indices …

[image] simeonschaub: If you do inds = collect(CartesianIndices(A)) , you can modify inds just like any other array, if that’s what you’re looking for. Maybe you can specify a bit more, what behavior exactly you want, so we can help you. julia> x=zeros(2); ind=collect(CartesianIndices(x)) 2-e…

You’ll see much better performance all around if you just use :s instead. You can construct the CartesianIndices you want, though: julia> CartesianIndices((axes(A)..., 1:1)) 10×1 CartesianIndices{2,Tuple{Base.OneTo{Int64},UnitRange{Int64}}}: CartesianIndex(1, 1) CartesianIndex(2, 1) ...

[image] mbauman: You’ll see much better performance all around if you just use :s instead. By :s, you mean metaprogramming? I tried the following: d=2 A_t = zeros(10,10,2) A = ones(10,10) eval(Meta.parse("A_t["*repeat(":,",d-1)*":,1] = A")) which works. However, when I embed the above in a …

No, I mean just using colons directly to select entire dimensions, like A_t[:, 1] or B[:, :, 1], etc. Those can be much more efficient than CartesianIndices because they convey additional meaning — we know (in the type domain) that they select an entire dimension whereas CartesianIndices may select…

Thanks! Both fill(:,d) and EllipsisNotation.jl are what I’m looking for. And the former is almost as fast as A_t[:,:,1] in 2D.

Indexing multidimensional arrays of arbitrary dimension

General Usage Performance

Topic		Replies	Views
Accessing multi-dimensional arrays in 1 go General Usage question , array	5	474	September 15, 2020
Iteration with CartesianIndices(array) is slow when the dimension of the array is large Performance indexing	2	840	August 25, 2019
Extra dimensions and CartesianIndices General Usage indexing	2	311	April 19, 2023
Cheapest way for creating a CartesianIndex General Usage question , cartesianindices	9	560	November 6, 2022
CartesianIndices syntax General Usage indexing	6	737	April 9, 2019

Indexing multidimensional arrays of arbitrary dimension

Related topics