[Mlir-commits] [mlir] [mlir][tosa] Fold PadOp to tensor operations (PR #132700)

[Georgios Pinitas]

GeorgeARM wrote:

There are two issues.

Pool operations require the padding to be less than the kernel size. We can handle this in the canonicalization pattern and avoid folding if this is not honoured.

When it comes to level checking we can't perform deterministically any check in the canonicalization.
This is a validation aspect pass.

When it comes to output sizes I don't see any issues.

So we can update the canonicalization but there is an edge case (extreme) that it could lead to non-compliant IR. Am fine with this. Otherwise the safe option is a controlled optional pass.

https://github.com/llvm/llvm-project/pull/132700