This should be a fairly straightforward translation of each output's mode to sparse CSC, then hstacking.