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

[Oleksandr Alex Zinenko]

ftynse wrote:

Another fly-by: I am generally in line with @rolfmorel later edits as a direction. Transform dialect was intended to make transforms more composable and initially replaced a monolithic "LinalgCodegenStrategy" object. Let's not replicate that object with transform dialect.

The current granularity of pattern exposure being low does not mean we can't have pattern groups exposed as well. Maybe we can consider having lowerings within transform dialect (the op for a pattern group lowers to a set of ops for individual patterns/small groups) or introduce a library/macro mechanism similar to how we have `transform.include` for named sequences.

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