TC-GNN: Bridging Sparse GNN Computation and Dense Tensor Cores on GPUs