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

[Luke Hutton]

lhutton1 wrote:

I realised that I never properly responded to the comments here regarding the level limitation. I feel the correct solution to this problem lies as part of a larger chunk of work which probably deserves a separate discussion. In the essence of moving forwards, perhaps we can implement your suggestion:
> Change the canonicalisation to work under the condition that it doesn't break the expectations of the operator

in the short term, such that folding only happens when the new pad values are known to be less than 8192 (MAX_KERNEL when `tosa_level=8k`)? This way we can guarantee no invalid IR is generated from a input that is valid


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