[mcabbott]

Two-argument show printed GPU arrays exactly like their CPU copy, so a GPU array inside a tuple, NamedTuple or struct looked like an Array:

julia> (x = CuArray([1.0]), y = cu([2.0 3.0]))
(x = [1.0], y = Float32[2.0 3.0])                    # before
(x = CuArray([1.0]), y = CuArray(Float32[2.0 3.0]))  # after

The output is a call to the storage type around the materialized CPU copy, so it evaluates to an equal array, wrapped arrays included:

julia> m = cu([1 2; 3 4]);

julia> repr(m')
"CuArray([1 3; 2 4])"

julia> repr(view(m, 1, :)), repr(similar(m, 0, 2)), repr(CuArray(fill(1.0)))
("CuArray([1, 2])", "CuArray(Matrix{Int64}(undef, 0, 2))", "CuArray(fill(1.0))")

Wrappers with a printer of their own keep it, e.g. on Julia 1.12+ Tridiagonal(CuArray([3.0]), CuArray([1.0, 2.0]), CuArray([4.0])) and adjoint(CuArray([1.0])). Like Base's eltype prefix, the call is left out when the context's :typeinfo is exactly the array's type, as for the elements of a concretely typed container:

julia> repr([m, m])
"CuArray{Int64, 2, CUDACore.DeviceMemory}[[1 2; 3 4], [1 2; 3 4]]"

julia> repr(Any[m])
"Any[CuArray([1 2; 3 4])]"

3-arg (REPL) display is unchanged, except where Base delegates it to 2-arg show: empty arrays in a :compact context, and on Julia 1.13+ empty vectors, which now display as e.g. CuArray(Int64[]) rather than Int64[].

The type parameters of the storage are not shown, so a CuArray in unified memory shows as CuArray(...) too. Zero-dimensional arrays used to throw a scalar-indexing error and now show as above. Wrapped arrays in a container typed on the wrapper ([m']) show as Adjoint{...}[CuArray(...)], which does not convert back; Base has the same limitation for CPU wrappers.

Implementation: show(::IO, ::AbstractGPUArray) handles unwrapped arrays and decides the :typeinfo elision, because Base replaces :typeinfo before calling its internal helpers. Wrapped arrays still go through Base's show and the existing _show_nonempty, _show_empty and show_vector overrides, so wrappers with their own show method keep it.

Supersedes JuliaGPU/CUDA.jl#1197.

[christiangnrd]

I fixed up this PR and it's pretty simple is this something we still want?

[maleadt]

Rebased and reworked, this should be good to go once CI passes.