Nested SubArray Scalar Indexing

[btmit]

v2.0.0 fixed a number of issues with subarrays, but nested calls still seem to have issues. I'm wondering if there is a way to "flatten" the different calls to view/reshape. See the small example below. In my application the different calls to view are on opposite sides of a function boundary so they can't be combined.

using CUDA
CUDA.allowscalar(false)

a = CUDA.zeros(4,3,2)
b = view(a, :,:,1)
c = CartesianIndex.([(1,3), (2,1)])
c = view(b, c)

ERROR: scalar getindex is disallowed

[maleadt]

Nested wrappers is JuliaGPU/Adapt.jl#21, so not supported.

We do support limited view nesting, but we rely on Base for that with a GPU version of DenseArray. And this view with Cartesian indices isn't a dense array. GPUArrays' copy methods are more generic, but those can only use Array because of #491.

[btmit]

Is there a workaround I could use to flatten this sort of thing? The following works, so it seems like there might be a way to reform the view from the parent on the second call?

using CUDA
CUDA.allowscalar(false)

a = CUDA.zeros(4,3,2)
c = CartesianIndex.([(1,3,1), (2,1,1)])
c = view(a, c)

[maleadt]

Not that I know, we're just reusing Base's view machinery here, and it doesn't look like it understands what you're doing enough to flatten the indices automatically.

[maleadt]

Works again after #498, now that we eagerly flatten some wrappers. But ultimately this remains Base not supporting these kinds of nested wrappers, and it just happens to work because of an optimization here, not because we fundamentally support this pattern.