Broadcast in Julia 1.11 requires getindex/setindex! to work on multiple integer indexing

[charleskawczynski]

This didn't used to be the case in Julia 1.10, and this new requirement means that we cannot reserve this indexing for special behavior.

For example, the fix needed: CliMA/ClimaCore.jl#2018

[charleskawczynski]

I'll try to make a reproducer at some point

[mbauman]

Interesting — these look like they're all non-AbstractArrays who have axes and previously supported get/setindex with CartesianIndex alone?

[charleskawczynski]

they're all non-AbstractArrays who have axes and previously supported get/setindex with CartesianIndex alone?

Yeah, that's accurate. I put a fair bit of work into removing multiple integer indexing. We'd like to reserve linear indexing (i.e., multiple integer indexing) so that we can special-case fixes for #28126 (including the GPU-variant related issue resolution efforts: JuliaGPU/GPUArrays.jl#454, JuliaGPU/GPUArrays.jl#464).

[charleskawczynski]

Is there anything more that I can do to help this get fixed? (or is there any way we can work around this?)

[mbauman]

I mean, it’s not abundantly obvious to me that broadcast should exclusively use one or the other — or which one in the first place. I certainly couldn’t have told you that broadcast always generated CartesianIndex-based getindexes on 1.10.

At a caller level, I have always implicitly assumed that the two should be perfectly equivalent. At the implementation level, AbstractArray considers n-ary Int to be the “canonical“ one. Having different behaviors between the two seems fraught for many algorithms, not just broadcast.

[charleskawczynski]

Ok, yes, I agree with you regarding design choices. We’re pretty heavily using broadcast in the CliMA org, and the 2x slowdown from Cartesian indexing is important enough that we need an alternative—whether that is some way to specialize on linear indexing, or roll our own custom version of broadcast.

I’ll look into the former, first, since that may be significantly less work.

[charleskawczynski]

It turns out that I was actually able to fix this with dispatch at a higher level.