use style dispatch in materialize

[vchuravy]

x-ref: JuliaLang/julia#35620

[maleadt]

What's the advantage of this? Tests?

[vchuravy]

dest .= expr, the destination can now be anything that promotes to the right broadcast style instead of being a fixed union.

[piever]

It seems like this is actually needed to get nice interoperability between StructArrays broadcast and GPU arrays broadcast, see JuliaArrays/StructArrays.jl#211 (comment). A StructArray whose component are GPU arrays is a good example of a suitable container that is not covered by the current copyto! rule.

This change (plus potentially moving the backend function to a lightweight package like Adapt.jl to avoid taking a dependency here) should allow external packages with custom broadcast (like StructArrays) to interoperate with GPUArrays broadcast.

[vchuravy]

@piever want to take a stab at this? As you can tell I run out if time to work on it shortly after opening the PR

[N5N3]

@vchuravy I guess the following is enough?

@inline function materialize!(::Style, dest, bc::Broadcasted) where {Style<:AbstractGPUArrayStyle}
    return _copyto!(dest, instantiate(Broadcasted{Style}(bc.f, bc.args, axes(dest))))
end

@inline Base.copyto!(dest::BroadcastGPUArray, bc::Broadcasted{Nothing}) = _copyto!(dest, bc) # Keep it for ArrayConflict

@inline Base.copyto!(dest::AbstractArray, bc::Broadcasted{<:AbstractGPUArrayStyle}) = _copyto!(dest, bc)

@inline function _copyto!(dest::AbstractArray, bc::Broadcasted)
    axes(dest) == axes(bc) || Broadcast.throwdm(axes(dest), axes(bc))
    isempty(dest) && return dest
    bc′ = Broadcast.preprocess(dest, bc)
    ...

Local test shows no problem. Is it OK to open a new PR?

[piever]

@N5N3, I think it would be great if you could open PRs here and on StructArrays to check that everything works. I confess I am a bit out of my depth on the technicalities of the broadcast machinery.

[maleadt]

Superseded by #393 IIUC