Implement convert between Array and CuArray like unsafe_wrap

[Jutho]

I am not sure if this qualifies as a bug or a feature request, but I was wondering whether the constructor CuArray{T,N,CUDA.HostMemory}(A::Array{T,N}) should unsafe_wrap(CuArray, A)?

The current general implementation at https://github.com/JuliaGPU/CUDA.jl/blob/master/src/array.jl#L404 will allocate new host memory and then copy the contents of A into this. Because of this, also convert(CuArray{T,N,CUDA.HostMemory}, A::Array{T,N}) (over at GPUArrays.jl) and ultimately adapt(CuArray{T,N,CUDA.HostMemory}, A::Array{T,N}) will result in new host memory being allocated. Hence, the only way to actually reuse A is via unsafe_wrap(CuArray, A), but that does only work for A::Array and not for any of the wrapper types that one could feed into adapt.

[Jutho]

On second thought, I assume Julia's array constructors are typically assumed to make a copy of the incoming data. But probably convert or adapt(CuArray{T,N,CUDA.HostMemory}, A::Array{T,N}) could be special cased to return unsafe_wrap(CuArray, A).

In the same way, adapt(Array, A::CuArray{T,N,<:Union{CUDA.HostMemory,CUDA.UnifiedMemory}) could be special cased to return unsafe_wrap(Array, A)?

[maleadt]

I would agree that convert could be made to work like this, while constructors probably should remain copying. However, there's a reason this operation is currently marked unsafe: We can't track the owning memory object, so when you e.g. unsafe_wrap an Array as a CuArray, and the Array is GC'd, the CuArray becomes invalid too. Without that, I don't think we can reasonably make convert call / behave like unsafe_wrap.

EDIT: Thinking about this some more, I'm not entirely convinced if convert or adapt should result in an aliased object. Are there precedents in Base or packages?