end-to-end GNN training and inference

[Dibbya40]

Hello. Is the code for getting end-to-end GNN training and inference time, available for us?

[Huyuwei]

FeatGraph is now partially integrated into DGL. Check https://github.com/dmlc/dgl/tree/master/featgraph

Some of the techniques in FeatGraph (e.g., feature dimension parallelization) have been used to optimize the hand-written kernels in DGL, while some techniques (mainly tiling) are not incorporated.

If you want to do end-to-end benchmarking, it should be ok to just run the master-branch DGL.

[Dibbya40]

Thank you @Huyuwei

[Dibbya40]

Hello. I had one more question. As FeatGraph, which is based on TVM, is partially integrated to DGL. Are the GCN's forward and backward computation code written in TVM?

Thank you.

[yzh119]

No DGL still uses CuSparse.

[Dibbya40]

Thank you @yzh119