view() returns CuArray instead of view for 1-D CuArrays

[omlins]

MWE:

using CUDA

nx = 3
B = CUDA.zeros(nx)
view(B, 1:nx-1)
typeof(view(B, 1:nx-1))

[maleadt]

view doesn't have to return a SubArray. This object shares memory with the parent, which is all that matters.

[omlins]

Thanks for looking into this, @maleadt . For multi-dimensional arrays, it creates a SubArray; for Base.Array, it it does in any case:

julia> using CUDA

julia> nx = ny = nz = 3;

julia> B = CUDA.zeros(nx);

julia> C = CUDA.zeros(nx,ny);

julia> D = CUDA.zeros(nx,ny,nz);

julia> E = zeros(nx);

julia> F = zeros(nx,ny);

julia> G = zeros(nx,ny,nz);

julia> typeof(view(B, 1:nx))
CuArray{Float32, 1, CUDA.DeviceMemory}

julia> typeof(view(C, 1:nx, 1:ny))
SubArray{Float32, 2, CuArray{Float32, 2, CUDA.DeviceMemory}, Tuple{UnitRange{Int64}, UnitRange{Int64}}, false}

julia> typeof(view(D, 1:nx, 1:ny, 1:nz))
SubArray{Float32, 3, CuArray{Float32, 3, CUDA.DeviceMemory}, Tuple{UnitRange{Int64}, UnitRange{Int64}, UnitRange{Int64}}, false}

julia> typeof(view(E, 1:nx))
SubArray{Float64, 1, Vector{Float64}, Tuple{UnitRange{Int64}}, true}

julia> typeof(view(F, 1:nx, 1:ny))
SubArray{Float64, 2, Matrix{Float64}, Tuple{UnitRange{Int64}, UnitRange{Int64}}, false}

julia> typeof(view(G, 1:nx, 1:ny, 1:nz))
SubArray{Float64, 3, Array{Float64, 3}, Tuple{UnitRange{Int64}, UnitRange{Int64}, UnitRange{Int64}}, false}

[maleadt]

That's correct. It tries to figure out if the view is contiguous. If it may be, a CuArray is returned.

Why is this a problem?

[omlins]

view doesn't have to return a SubArray. This object shares memory with the parent, which is all that matters.

Wouldn't it be more helpful for the programmer if it returns always a SubArray for views on CuArrays? Also, that's the behavior for Base.Arrays and therefore in order to have maximum consistency, I would have expected the same behavior. Furthermore, if the view returns simply a CuArray that shares the memory, I cannot conveniently access the parent, indices,... - i.e. the fields provided by a SubArray - if i need to do so. In my use case, I would like to access the parent etc...

[maleadt]

I get what you're asking and why, but this won't change. There is no contract for these calls to return a SubArray, and the very real problems with wrapped arrays are much more problematic than the inconveniences resulting from this design. From the top of my head: much more complex subtype queries, notably in linalg wrappers, leading to significant package load time regressions; many more ambiguities, again especially around the linalg stuff; triggering generic fallbacks where we already dispatch on wrapped arrays (since we currently do not support multiply-wrapped arrays); etc. If anything, this first needs some fixes in Base, e.g., JuliaLang/julia#51910, before we can revisit this.

[omlins]

@maleadt Okay, I understand. Thanks for clarifying the situation!