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

[Andrzej Warzyński]

banach-space wrote:

> I can't speak for the other Ops, but the "channel first" ones I've been adding are in order to support lowering torch IR to linalg without additional transpositions in torch-mlir.

Avoiding unnecessary transposition is good, but will this convolution get vectorized? If not, does this make any difference? Without vectorization, things will be slow _with_ and _without_ transposing.

Just wanted to draw people's attention to the fact that adding a new named Op, especially a convolution, is just one step towards good lowering 😅 And  that there are some non-trivial gaps in how convolutions are vectorized.

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