add a note about reinterpret's memory layout

[johnnychen94]

I didn't add the StructArray{Point{Float64}}(X, dims=2) trick here because it is still quite strange to me; it has a strong assumption to how you interpret the data from the raw contents, but the confusion I get from #197 is how to get the "actual" memory layout from an already constructed StructArray.

closes #197

[piever]

Thanks, I've added some changes to wording, and a possible couple of sentences to put at the beginning for context.

I didn't add the StructArray{Point{Float64}}(X, dims=2) trick here because it is still quite strange to me; it has a strong assumption to how you interpret the data from the raw contents

I actually think it should be mentioned in this section, because it is the easiest way to get a StructArray from a higher-dimensional array of primitive types. Some of the confusion, IMO, comes from the fact that, unlike Vector{ComplexF64}, in StructArray{ComplexF64} there is no single "block of memory" to speak of, but two separate vectors which will be at distinct locations. The only way to associate it to a single "block of memory" is to choose those vectors as contiguous views of an existing matrix.

As a meta comment, I think this section should go among the "Advanced" sections at the end of the README. reinterpret is a bit on the technical side in my opinion.

[johnnychen94]

If I use v = [1 3; 2 -1] then I get

julia> StructArray{ComplexF64}(v, dims=1)
0-element StructArray(StructArray(), StructArray()) with eltype ComplexF64 with indices 1:0

Is this expected?

[piever]

Are you sure? I get

julia> v = Float64[1 3; 2 -1]
2×2 Matrix{Float64}:
 1.0   3.0
 2.0  -1.0

julia> StructArray{ComplexF64}(v, dims=1)
2-element StructArray(view(::Matrix{Float64}, 1, :), view(::Matrix{Float64}, 2, :)) with eltype ComplexF64:
 1.0 + 2.0im
 3.0 - 1.0im

which is the expected behavior. Btw, dims=ndims(v) would be the way to get contiguous views as component arrays (selecting on the last dimension).

[johnnychen94]

Sorry , I mean:

julia> v = Int[1 3; 2 -1]
2×2 Matrix{Int64}:
 1   3
 2  -1

julia> StructArray{ComplexF64}(v, dims=1)
0-element StructArray(StructArray(), StructArray()) with eltype ComplexF64 with indices 1:0

julia> StructArray{ComplexF64}(reinterpret(Float64, v), dims=1)
2-element StructArray(view(reinterpret(Float64, ::Matrix{Int64}), 1, :), view(reinterpret(Float64, ::Matrix{Int64}), 2, :)) with eltype ComplexF64:
 5.0e-324 + 1.0e-323im
 1.5e-323 + NaN*im

julia> v = ComplexF64[1 3; 2 -1]
2×2 Matrix{ComplexF64}:
 1.0+0.0im   3.0+0.0im
 2.0+0.0im  -1.0+0.0im

julia> StructArray{ComplexF64}(v, dims=1)
2-element view(::Matrix{ComplexF64}, 1, :) with eltype ComplexF64:
 1.0 + 0.0im
 3.0 + 0.0im

julia> StructArray{ComplexF64}(v, dims=2)
2-element view(::Matrix{ComplexF64}, :, 1) with eltype ComplexF64:
 1.0 + 0.0im
 2.0 + 0.0im

Interpreting the output is quite, hmmm, unintuitive. Maybe I just hit some undefined behaviors.

[piever]

Ah, I see, it behaves a bit funny if the types don't match. It's probably because it has some logic to support nested cases, I've opened #200 to track this. Note that this constructor does not allocate, it can't use a matrix of integers to store components of ComplexF64.

[piever]

Actually, here and below the memory is always v, this constructor does not allocate.

[johnnychen94]

Oh thanks for pointing this out! I didn't realize this, I only did a quick check @btime StructArray{ComplexF64}(v, dims=1) with small v = Float64[1 3; 2 -1] and I thought the memory allocations means copy 😂

julia> @btime StructArray{ComplexF64}(v, dims=1);
521.754 ns (8 allocations: 352 bytes)

[piever]

Maybe best to just mention the actual caveat, that is that v must be typed correctly. Rather than memory layout (as the only in memory object is always v), using dims=ndims(v) is done to get the best performance, because then the components of the StructArray are contiguous views (i.e., things like @view v[:,1], which is the most efficient in column-major languages).

[johnnychen94]

Getting a better understanding of this now. Thanks for the feedback!

Hope this is the last commit 😄

[piever]

Great, I'm glad doing this was instructive!