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

[Stella Laurenzo]

stellaraccident wrote:

Couple of comments: 

1. I think the most widely used op was probably the matmuls, and we just got that reconverged (still some cleanup but the main thrust resolved).
2. Convolution naturally has more variants and practical decisions need to be made. I assume the answer is not one op. But neither is it ten.
3. It is a historic anomaly that channels last got the priority when the most dominant framework operates channels first. We can't block some amount of progress on fixing that bug, regardless of POR. It would be good to treat these as high cost exceptions and only do it with a strong expressed need, though... Since it adds to the work needed to fix the bigger issue.
4. We should publish a plan of record on where convolution needs to get to, informed by the matmul work. It's unfair to just tell people no on such an important case without so much as a written plan on what to do about it.

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