Continuum indices about continuumarrays.jl HOT 10 OPEN

This may be related.
#17

It seems that only linear continuum indices is supported.

from continuumarrays.jl.

The getindex approach is a workaround.

julia> g=getindex.((f,),f)
getindex.(Tuple{QuasiArrays.BroadcastQuasiVector{Float64, typeof(Base.literal_pow), Tuple{Base.RefValue{typeof(^)}, Inclusion{Float64, IntervalSets.ClosedInterval{Int64}}, Base.RefValue{Val{2}}}}}, (Inclusion(0..1)) .^ 2)

julia> g[0.5]
0.0625

So we can do this:

julia> Base.getindex(A::AbstractQuasiArray,B::AbstractQuasiArray) = getindex.((A,),B);
       g=f[f];
       g[0.5]
0.0625

I'm not sure about the efficiency.

from continuumarrays.jl.

It looks like the issue is bounds-checking. That is, we want issubset(Inclusion(0..1).^2, Inclusion(0..1)) to return true.

We could probably do something like:

issubset(a::BroadcastQuasiVector{<:Any, typeof(Base.literal_pow), <:Tuple{Base.RefValue{typeof(^)}, Any, Base.RefValue{Val{2}}}}, b::Inclusion{<:Any,<:ClosedInterval}) = issubset(a.args[1], Inclusion(sqrt(max(0,first(b)), sqrt(max(0,last(b))))

from continuumarrays.jl.

Did that "affine transform" issue use the same workflow? If so, we may cover a few other basic functions.

from continuumarrays.jl.

Actually, we pass through a call to union, which is what we use to determine the "range":

from continuumarrays.jl.

Adding a subtype of Map should solve my problem. However I would like generic continuum indices to work as well. One reason is that the expression f[f] is easy to understand. Another reason is that creating and maintaining interfaces for all known maps is exhaustive.

from continuumarrays.jl.

Agreed. Try my suggestion?

from continuumarrays.jl.

Do you mean the issubset or the union?

For generic quasivector, the range is usually hard to tell. I prefer passing the responsibility of boundschecking to users.

Although an unknown interval which may exceed bounds makes the whole thing bad-defined, in practice, we always evaluate a finite number of values and all variables reduce to finite vectors which can be boundschecked as usual.

Checking bounds for the whole interval does reduce the cost of checking every value later. I agree we can have this feature for common functions.

My thought is to let issubset to return true as default and override it for specific cases. I guess unsafe_view on continuumarrays won't cause memory problems.

from continuumarrays.jl.

What's the disadvantage of unsafe_view?

from continuumarrays.jl.

Right, Base.unsafe_getindex or unsafe_view might be a good interface

from continuumarrays.jl.