How to cut a sparse gpuArray

I am trying to save memory on my GPU by saving sparse data, but matlab gives this error whenever I try to cut a sparse gpuArray in the following way.

Error using gpuArray/subsref
Sparse gpuArrays are not supported for this function.

Ex.

A=gpuArray(sparse(randi([0,1],2000,2000)));

B=A(1:50,:);

Is there any way to cut sparse gpuArrays, or am I stuck transferring data between my CPU and GPU?

NOTE: This is a piece of example code. Although this piece of code will not have memory benefits from a sparse array, my application will.

Answer 1

[Matt J]

Here's a workaround equivalent to B=A(1:50,:), but I do agree it seems unnecessarily awkward.

B = gpuArray.speye(50,2000) * A;

Answer 2

[Sean de Wolski]

For the sparsity level of that matrix ~50%, the sparse array is much bigger than a regular array:

x = randi(1,100,100);
xs = sparse(x);
whos x xs
  Name        Size              Bytes  Class     Attributes

    x         100x100             80000  double              
    xs        100x100            160808  double    sparse

You're better off just using a regular array. And you can generate randi directly on gpu so it doesn't need to be transferred:

gpuArray.randi(1,100,100)

[Kevin Bender]

This is just an example, not the actual code. In my actual code, I am storing a sparse adjacency matrix that has less than 5% nonzero elements.