BlockGNN: Towards Efficient GNN Acceleration Using Block-Circulant Weight Matrices