[Mlir-commits] [mlir] [linalg] Add quantized version of `conv_3d_ncdhw_fcdhw` (PR #113953)

[Andrzej Warzyński]

banach-space wrote:

I share Renato's concerns and agree with his perspective.

As I understand, we're aligned on the need to reduce the number of named operations in Linalg, and there’s already ongoing work in this direction. This PR, while aiming to support a specific downstream target, seems to move us away from that goal and, therefore, feels counterproductive.

Supporting downstream targets is certainly valuable, and I recognize the benefit of this approach for your project:

> I use Linalg graphs lowered from Torch IR as a starting point for compilation to a custom dataflow accelerator.

However, from a pragmatic and long-term perspective, we should carefully consider the potential consequences. The risk is that downstream targets might get discontinued, or future downstream projects may have different needs. This could lead to even more named Ops, which would be challenging to maintain and reconcile.

I appreciate the work you've put into this PR, and I don’t want to discourage contributions, but I believe it would be beneficial to explore alternative approaches that align with our broader goals.

https://github.com/llvm/llvm-project/pull/113953