Missing dispatch for indexing of reshaped arrays

[ChrisRackauckas]

x = cu(rand(8))
W = @view p[reshape(1:4,2,2)]
W' # fails

This is because W is a SubArray`. However, the other method works:

x = cu(rand(8))
W = reshape(@view(p[1:4]),2,2)
W'

This is different from Array semantics:

x = rand(8)
W = reshape(@view(x[1:4]),2,2)
W'

x = rand(8)
W = @view x[reshape(1:4,2,2)]
W'

where the latter is more optimized with BLAS due to some of @mbauman 's fanciness.

[maleadt]

What do you expect here exactly? @view x[reshape(1:4,2,2)] is a non-contiguous SubArray, both on the CPU and the GPU, because of the indices. Calling transpose on that falls back to the generic implementation on the CPU, and can't possibly work on the GPU because of JuliaGPU/Adapt.jl#21 once again.

[ChrisRackauckas]

It's contiguous:

julia> x = reshape(1:100,10,10)
10×10 reshape(::UnitRange{Int64}, 10, 10) with eltype Int64:
  1  11  21  31  41  51  61  71  81   91
  2  12  22  32  42  52  62  72  82   92
  3  13  23  33  43  53  63  73  83   93
  4  14  24  34  44  54  64  74  84   94
  ⋮                   ⋮
  7  17  27  37  47  57  67  77  87   97
  8  18  28  38  48  58  68  78  88   98
  9  19  29  39  49  59  69  79  89   99
 10  20  30  40  50  60  70  80  90  100

julia> @view x[reshape(1:4,2,2)]
2×2 view(::UnitRange{Int64}, [1 3; 2 4]) with eltype Int64:
 1  3
 2  4

[maleadt]

It may be in practice, but view with indices like that cannot be detected as such:

julia> W = @view x[reshape(1:4,2,2)]
2×2 view(::UnitRange{Int64}, [1 3; 2 4]) with eltype Int64:
 1  3
 2  4

julia> ans isa Base.FastContiguousSubArray
false

vs

julia> W = @view x[1:4]
4-element view(::UnitRange{Int64}, 1:4) with eltype Int64:
 1
 2
 3
 4

julia> ans isa Base.FastContiguousSubArray
true

[ChrisRackauckas]

I think I can work around it: SciML/DiffEqFlux.jl#451

[ChrisRackauckas]

Note that on CPUs, that other form is treated as a contiguous array by BLAS via JuliaLang/julia#34729

[maleadt]

That's a good reference, thanks. I'm wary to copy it here though, because these unions have a tendency of wrecking package import latency...

[maleadt]

The original MWE here seems to work now.