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

[Renato Golin]

rengolin wrote:

> but it does feel like the abstract goal of "avoiding unnecessary transpositions" is a general good, even though some backends and lowering paths might only benefit minimally.

This is a false correlation. Yes, avoiding unnecessary transposition is a good thing, but the way you're achieving is by bloating the namespace beyond usefulness. This is the same problem we have with `matmul_transpose_a` and `matmul_transpose_b` etc. It does not actually _solve_ the problem, just moves it somewhere else.

See https://github.com/llvm/llvm-project/pull/104783 for a path to a proper solution.


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