TC-GNN

TL;DR: This paper propose to co-densing (SpArch also propose co-densing) the sparse matrix $$A$$ in SpMM $$C=A\times B$$ to reduce the fragmentation when using Tensor Cores for SpMM computations on GPU.

My Digest

The co-densing could be described as a special format rewrite rules in SparseTIR, worth trying it and add an example in the documentation.

Email: expye@outlook.com

Date: 2022-08-17 Wed 00:00