Slicing an array creates a copy instead of a reference

[raichu]

I'm in the middle of a migration from C++ and I'm having a hard time in understanding Julia "slices".

I have a vector a. When I do b=a[1:2] and b[1] = 10 for example, a is not affected.
That is, the "slicing" operation a[1:2] creates a copy instead of working with the original data.

Is this a bug or intended behavior?

If it is intended behavior, what is the correct way of getting a reference to a piece of the original vector? (whose type is still Vector, not SubArray; I have functions with explicit type annotations for type safety).

[nalimilan]

This is the intended behavior, but it's probably going to change. See #5556.

For now, I think you need to use SubArrays. You could change your annotations to check for AbstratArray rather than Array: this doesn't remove much safety and makes your code more generic.

[raichu]

Well, that's disappointing.

[raichu]

How would it work exactly?
Apparently, slice and slicedim returns a copy as well.

Here's what I'm trying to do; I'm using NLopt.jl and it's API requires all the data is given as a big Vector (in my case it's of size 5_N_N). The data is actually made of 5 NxN Float64 matrices. So I need to find a way of re-interpreting the original Vector in terms of five N*N matrices. And I should be able to work on the original data (to work with NLopt's C code), not a copy of the data.

Is there a way of achieving this in Julia today?

[raichu]

Just tried reshape(sub(v, 1:N*N), N,N). This interestingly gives a copy of the array.
Using only sub or only reshape, I get a reference to the original array, but when used together, they give a copy of the data instead.
Is this not a bug?

[nalimilan]

You can simply call reshape on the vector to make it a 3-d array, and then use sub to extract individual matrices from it.

[raichu]

Sorry, I'm not that familiar with Julia, what is a 3d array? How do I do that exactly? reshape(v, 5, N ,N) doesn't seem to be giving something pretty. And are you sure that it's not going to be a copy of the original data? (see my previous comment).

Also, doesn't sub give an array instead of a matrix?

[raichu]

Can you give me a working code?
If it is such a simple thing that can be done just by reshape and sub, it should be a one-liner, right?

[nalimilan]

Try this, it doesn't copy the arrays:

# Vector you get from NLopt (here with 2 5x5 matrices)
julia> a = rand(2 * 25)
# 3-dimensional array holding your 5x5 matrices
julia> b = reshape(a, (5, 5, 2))
# Extract the first 5x5 matrix
julia> c = sub(b, :, :, 1)

EDIT: fix the order of dimensions.

[raichu]

Thanks, but it gives 2x5 matrix instead of a 5x5 matrix.

Chaning (2, 5, 5)->(5, 5,2) did the job.

[raichu]

Just another quick question; the type of the result is
5x5 SubArray{Int64,2,Array{Int64,3},(UnitRange{Int64},UnitRange{Int64},Int64)}

Would there be any performance penalties with this abstraction? Or is the optimizer clever enough to optimize this thing as a normal, contiguous array?

[ivarne]

Sub arrays are not known for their performance. I think the ArrayViews from #5556, is likely to have better performance.

[kmsquire]

Duplicate of #3701.

[lindahua]

Please check out ArrayViews package at this point, which provides what you want (a performant non-copying view).

[JeffBezanson]

If you need to be sure data is referenced in-place, you can hack around with pointer and pointer_to_array. Just make sure there is a reference to the original data object somewhere.
Closing as dup of #3701.