copy! methods for views with linear indices

[samuelpowell]

I have an application in which an (expensive) linear operator is implemented on the GPU. Methods such as GMRES from IterativeSolvers attempt to apply the operator to a RHS which is a view of another larger array. Since there are no copy! methods for views, a naive implementation fails.

In the case of GMRES (and one would imagine most performance critical applications), the views will be an appropriate sized column of a matrix, thus the memory is linear and the copy operation could presumably be performed without too much pain.

I am not that familiar with the SubArray architecture - would it be to implement a copy! method which accepts a view, so long as the indices are linear and the dimensions correct?

[samuelpowell]

Having looked a little at the documentation, perhaps it would be possible simply to add a new method with signature

function Base.$copyfun(dst::CuArray{T}, src::SubArray{T,N,A,I,true}) where T

which checks for unit stride, and then continues in the same vein as the standard copy methods. It appears in that in this case passing a pointer and length to the CUDA memcpy function should suffice.