Missing indexing on GPUSparseDeviceMatrixCSC #730

The lack of specialized getindex methods for GPUSparseDeviceMatrixCSC makes it so that we cannot index into it (the previous CuSparseDeviceMatrixCSC could). As a consequence, we also cannot index into GPUSparseDeviceColumnView. I understand that it is not the most efficient as we need to perform a binary search, but in my use case, it is the only option.

MWE in terms of GPUSparseDeviceColumnView

using CUDA, SparseArrays

A = sprand(10, 10, 0.4)
A_gpu = cu(A)
a_gpu = @view A_gpu[:, 1]

function kernel(out, a)
   acc = 0.0
   i = 1
   while i <= length(a)
       acc += a[i]
       i += 1
   end
   
   out[1] = acc
   return nothing
end

out = CUDA.zeros(1)

@cuda kernel(out, a_gpu)

The kernel can equivalently be written as the following (subarray.jl:313).

function kernel(out, a)
   acc = 0.0
   i = 1
   while i <= length(a)
       index = Base.reindex(a.indices, (i,))
       v = a.parent[index...]
       acc += v
       i += 1
   end
   
   out[1] = acc
   return nothing
end

In this case, it succeeds with reindexing, i.e. to find the index within the parent (A_gpu), but since GPUSparseDeviceMatrixCSC lacks appropriate indexing methods, it fails with the error message (under julia with option -g2).

index: (1, 1)
ERROR: a exception was thrown during kernel execution on thread (1, 1, 1) in block (1, 1, 1).
Stacktrace:
 [1] error_if_canonical_getindex at ./abstractarray.jl:1357
 [2] getindex at ./abstractarray.jl:1341
 [3] kernel at ./REPL[7]:6