When you create a sparse matrix in for loop with OpenCL backend, the matrix is always the first one, i.e. it doesn't get updated. On CUDA and CPU backend everything works normally. The following minimal example demonstrates this:

array output;
float value[] = { 1, 1, 2, 2 };
int index[] = { -1, 1, 2 };
int row[] = { 0, 2, 2, 0, 0, 2 };
int col[] = { 0, 1, 0, 1 };
array values(4, 1, value, afHost);
array rows(6, 1, row, afHost);
array cols(4, 1, col, afHost);
array S;

for (int i = 0; i < 2; i++) {

	S = sparse(2, 2, values(seq(index[i] + 1, index[i + 1])), rows(seq(3 * i, 3 * (i + 1) - 1)), cols(seq(index[i] + 1, index[i + 1])));
	output = dense(S);

	af_print(output);


}

af:: info OpenCL:
ArrayFire v3.5.0 (OpenCL, 64-bit Windows, build 05999f3)
[0] NVIDIA: GeForce GTX TITAN, 6144 MB

CUDA:
ArrayFire v3.5.0 (CUDA, 64-bit Windows, build 05999f3)
Platform: CUDA Toolkit 8, Driver: CUDA Driver Version: 9000
[0] GeForce GTX TITAN, 6144 MB, CUDA Compute 3.5