Michael Abbott

But if I set `V2[2] = 99` then `V` is altered. However `typeof(V2)` and `dump(V2)` show no evidence of this, so perhaps at some level it's ignorant of the connection,...

BTW, I just saw https://github.com/JuliaLang/julia/pull/30973 and wonder if that issue (28126) might be what is going on here. Haven't tried it out yet.

I remain puzzled by the `V2 = reshape(V, 1,:)` issue, although it's not obviously this package's fault.

I do think it's the intended behaviour, although I don't see it mentioned in the manual. Maybe broadcast never produces a 0-array, e.g. `sqrt.(fill(pi)) isa Float64`.

https://github.com/JuliaLang/julia/issues/28866 is the issue, it seems.

One possible fairly straightforward rule would be drop all special indices along the direction of concatenation, and keep them when orthogonal. Thus `hcat(zeros(-1:1), ones(-1:1))` would have axes `(-1:1, OneTo(2))` (and...

The present behaviour of making `:` always start from 1 struck me as sensible. Here are some examples where the proposed rule (if I understood right) would give offsets that...

Perhaps of interest, I had a go at making something like this work, and now it's registered: ```julia using NamedPlus A, B = ones(x=10, t=15), ones(x=10, y=20); @named C{x,t,y} =...

I think AxisArrays never got its broadcasting updated for Julia 1.0, the old machinery was much simpler (but less powerful). It could surely be done though. My `@named` macro only...

Note that the particular function here is `NNlib.batched_mul`, which goes to a CUDA kernel. And in fact also `TensorCore.boxdot`, which just reshapes & calls `*`. One may be faster than...