Space usage of nested tuples

General Usage
1

Are vector elements (eg tuples with three fields each) in a structure such as myList = Vector{Tuple{UInt32, UInt16, UInt16}} stored as pointers to tuples, or directly as three UInts? In other words, would myList and myMatrix (initialzed as zeros(N, 3) ) take the same amount of space on a per element basis?

2

Directly (“inline”) as three integers, much like an array of struct. In general, immutable “bitstype” data is stored inline in an array (or elsewhere), without pointers.

Note that a mixed-type tuple will be subject to struct padding, so there may internally be additional padding bytes between entries depending on the sizes. In the particular case of Tuple{UInt32, UInt16, UInt16} there is no padding and it requires 8 bytes per entry:

julia> a = Tuple{UInt32, UInt16, UInt16}[(3,4,5)]
1-element Vector{Tuple{UInt32, UInt16, UInt16}}:
 (0x00000003, 0x0004, 0x0005)

julia> sizeof(a)
8

but for example Tuple{UInt8, UInt64} occupies 16 bytes since the second entry is padded to an 8-byte boundary:

julia> a = Tuple{UInt8, UInt64}[(3,4)]
1-element Vector{Tuple{UInt8, UInt64}}:
 (0x03, 0x0000000000000004)

julia> sizeof(a)
16
1 Like
3

Yes. ismutabletype(Tuple{UInt32, UInt16, UInt16}) == false and most immutable objects are stored inline in arrays.

As for memory use, I will remark that the two aren’t quite comparable. Your tuples are heterogeneous while a Matrix must be homogeneous. But for homogeneous tuples, the two would be comparable except for possible struct padding (as discussed in the above post).

If you’re considering a layout like zeros(UInt32,N,3), you might also consider the StaticArrays.jl package for forming a Vector{SVector{3,UInt32}}. This is memory-wise identical to a Vector{NTuple{3,UInt32}} (or a zeros(UInt32,3,N) assuming no padding – note the dimensions were swapped from your question as Julia is column-major) but provides array-like behavior from the SVectors rather than tuple-like behavior.

2 Likes
4

Thanks!