For example:

Asubset = A.sel[[1, 3, 5], :].new()

Asubset is the same shape as A, but only has values in rows 1, 3, and 5. The recipe for this example could be:

v = Vector.from_values([1, 3, 5], True, size=A.nrows)
Asubset = binary.second(v & A).new()

or

Asubset = Matrix(A.dtype, A.nrows, A.ncols)
Asubset[[1, 3, 5], :] = A[[1, 3, 5], :].new()

It's worth noting that binary.second(v & A) is also a recipe (one that I find quite useful!). I actually don't know which recipe is better or whether one recipe dominates the other.

I've actually encountered this operation many times, and I often find myself unsure about how I should do it, hence this proposal to have nicer syntax that lets us--not the user--decide how to do it.

This is of course low priority and non-essential. There are also questions to consider:

• Can this be used to assign where the object on the right is the same size as the object on the left?
• Can masks be used?

I don't recall ever needing assignment, so I would skip assignment for now and only do extract. Masking when extracting seems natural and can be done efficiently.

Do we like A.sel[...], or are there better ways to spell this?

I'm fine letting this issue sit for a while. There are other fancy indexing people may want to do, such as:

• index using the structure or values of a vector, A[v, :]
  • what is the output shape? Same as A? Using v.nvals feels weird.
• index using vector masks, A[v.S, w.V]
  • what is the output shape? Same as A?
  • could we mix styles, such as A[v.S, [1, 2, 3]]?
• index using tuples, such as A[((1, 5), (2, 10), (3, 15))] or A.points[(1, 5), (2, 10), (3, 15)]
  • this return a length 3 vector with elements from (1, 5), (2, 10), (3, 15)
  • or maybe this would be A.points[[1, 2, 3], [5, 10, 15]] for faster recipe and possible adoption in C

We can discuss other indexing possibilities in this issue, but if anybody has a need for one of these, please open another issue to request it.