similar(reinterpret(T, view(::CuArray, ...))) produces Array instead of CuArray

[simeonschaub]

It seems like this falls back to the generic implementation for AbstractArrays in Base, which always produces an Array:

julia> similar(reinterpret(SVector{4,Float32}, view(cu(rand(Float32, 10, 2)), 1:4, :)))
1×2 Array{SArray{Tuple{4},Float32,1,4},2}:
 [2.86328f16, 4.5768f-41, 0.0, 0.0]  [2.93855f16, 4.5768f-41, 2.86328f16, 4.5768f-41]

julia> similar(reinterpret(SVector{4,Float32}, cu(rand(Float32, 4, 2))))
1×2 CuArray{SArray{Tuple{4},Float32,1,4},2}:
 [0.683944, 0.111384, 0.644627, 0.645665]  [0.449726, 0.553998, 0.548373, 0.354193]

I think returning a CuArray would make more sense here, so generic code can still run on the GPU.

This is using CUDA 2.3.0 on Julia 1.5.3

[simeonschaub]

Alternatively, maybe this should be fixed upstream, so similar(reinterpret(T, ::A), ...) falls back to similar(::A, T, ...)?

[simeonschaub]

This should be fixed on Julia nightly, but would it be a good idea to try to fix this for CuArrays for older versions as well?

[maleadt]

We don't use ReinterpretArray anymore after #498, could you verify this works on master as expected?

[maleadt]

So this works with contiguous reinterpred views

julia> similar(reinterpret(Int32, view(CUDA.rand(2,2), :, 1)))
2-element CuArray{Int32, 1}:
 0
 0

julia> similar(reinterpret(Int32, view(CUDA.rand(2,2), 1, :)))
2-element Vector{Int32}:
 15
  0

[simeonschaub]

Awesome thanks! Works for me now.