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

[Renato Golin]

rengolin wrote:

> I dont know where these uses are coming from, but I'd like to just say that addition of new convolution ops is kind of getting out-of-hand. We have a lot and it doesnt cover all use cases, but it seems like might be time to move convolutions out of ODS generated named ops.

I second that. While we're not directly interested in convolution ops (so won't have time to move them to ODS ourselves), I strongly suggest those interested do so instead of adding yet-another op in the large soup of slightly different semantics.

There has to be a way to common them all up using ODS + C++ code that OpDSL just couldn't. This is the reason we moved `matmul` to ODS and we're much better for it.

For the record, the medium-term plan is to move all relevant Linalg operations to ODS and remove OpDSL for good, the code merged in this PR will not live long. So if you care about these operations, please come up with a reasonable (parametrized) implementation in ODS and move the rest there, too.

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