Input-sensitive dense-sparse primitive compositions for GNN acceleration