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

[Luke Hutton]

https://github.com/lhutton1 commented:

Thanks, the changes look great! My only concern would be when thinking about the limits for the padding for each operation:
- tosa.pad - requires `padding` dims to be of size governed by `shape_t` -  `MAX_LOG2_SIZE`
- tosa.conv2d/tosa.avg_pool2d/etc - requires `pad` attribute values to be of size `MAX_KERNEL`

Therefore, I believe it's possible (yet unlikely!) we end up folding two operations that are tosa compatible into an incompatible operation depending on the chosen tosa level. Because of this, do you think it would make sense to have this as a separate optional pass? We could start by always applying the transformation, similar to what is done here, but we can extend to allow users set the level they wish to conform to and let the pass make the decision about whether to fold in the future?

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