[Mlir-commits] [mlir] [mlir][linalg] Add mixed precision folding pattern in vectorize_children_and_apply_patterns TD Op (PR #148684)

[Rolf Morel]

rolfmorel wrote:

Here's my drive-by suggestion (which could be done in a follow-up PR as there isn't really anything wrong with the current PR, seeing as it keeps to an existing way of adding more patterns):

Merge `transform.structured.vectorize_children_and_apply_patterns` into `transform.structured.vectorize` modifying the latter to have (something like) the following syntax:

```
`transform.structured.vectorize`
    (`children_of`)?
    $target
    (`vectorize_nd_extract`)?
    (`flatten_1d_depthwise_conv`)?
    oilist(`vector_sizes` custom<DynamicIndexList>($vector_sizes, $static_vector_sizes, $scalable_sizes)) 
    (`with_patterns` $with_patterns)? attr-dict `:` functional-type(operands, results)
```

where we could choose that the `children_of` UnitAttr cannot be specified alongside `vector_sizes`.
`$with_patterns` would be an optional region where users could hook in their (TD-exposed) patterns, so as to make sure they run in the same rewriting fixpoint. In case we want to have that certain patterns are on by default (e.g. `transform.apply_patterns.vector.multi_reduction_to_contract` and `transform.apply_patterns.vector.multi_reduction_to_contract_patterns`) we could even go so far as to add

```
   (`without_patterns` $without_patterns)?
```

where you specify that these certain "without" patterns _shouldn't_ be added.

Just my two cents.

https://github.com/llvm/llvm-project/pull/148684