Heterogeneous Manycore Architectures Enabled by Processing-in-Memory for Deep Learning: From CNNs to GNNs: (ICCAD Special Session Paper).