Efficient computation of sparse approximate inverses