[Mlir-commits] [mlir] [mlir][sparse] Reuse input sparsity for materialized outputs (PR #215579)

[Aart Bik]

aartbik wrote:

Thanks for giving the additional background. But this is also exactly why I felt uncomfortable. You are adding a "fast-path" for just CSR into a sparse compiler that should work on generic sparse tensor formats. You are probably fixing a very specific performance issue for a very particular project. The "Initial Scope" mentions this restriction on the materialization, but makes no plans for generalizing this. I would much rather accept something that works for all sparse tensors, in the line of a sparse compiler.

@PeimingLiu , do you concur?

https://github.com/llvm/llvm-project/pull/215579