[Mlir-commits] [mlir] [mlir][sparse] support sparsifying 2:4 block sparsity (PR #71749)

[Yin Tong]

Ag-Cu wrote:

Thanks, This looks like an awsome job! I think I should to take a closer look.

By the way, I read a paper named [VENOM: A Vectorized N:M Format for Unleashing the Power of Sparse Tensor Cores](https://arxiv.org/abs/2310.02065), it is about an open source sparse library for any N:M Format instead of 2:4, and it can still make use of Sparse Tensor Cores. Do you think it is necessary to support N:M Format instead of only NV_24?(Due to current hardware limitations, this matter does not seem to be urgent? I think...)

We are trying to optimize this library, and later we are interested in MLIR's codegen to call Sparse Tensor Cores. If N:M Format is necessary, we may consider adding it to the work plan. And is there any information here that can help develop codegen?

Thanks for your time.

https://github.com/llvm/llvm-project/pull/71749