tree 56a8d689ecefa37f48113dad71220671d888a329
parent 507c7164ecc8ea52b21fc968147f8dd470dc484a
author Fabian Mora <fmora.dev@gmail.com> 1750326464 -0400
committer Copybara-Service <copybara-worker@google.com> 1750326710 -0700

[mlir][linalg] Add option to pad dynamic dims to `linalg::rewriteAsPaddedOp` (#144354)

This patch makes the following changes:

- Add a `ValueRange typeDynDims` argument to
`linalg::makeComposedPadHighOp`, allowing to pad a tensor with dynamic
dimensions using `tensor::createPadHighOp`.

- Add a `DenseMap<std::pair<unsigned, unsigned>, OpFoldResult>
sizeToPadTo;` option to `LinalgPaddingOptions`. This option allows
setting the size to use when padding a dimension of an operand, allowing
to pad operands even in the case they don't have a constant upper
bounding box. If the value is not provided, then the constant upper
bound is used by default.

- Add a `use_prescribed_tensor_shapes` option to
`transform.structured.pad`. If set to true then `tensor.dim` will be
used as dimensions to compute the size of the padded dim instead of
computing the constant upper bound.

- This patch also changes the behavior for computing the padded shape
`linalg::rewriteAsPaddedOp`, by using the newly added options in
`LinalgPaddingOptions`.

- Finally it adds tests verifying the behavior.

GitOrigin-RevId: 97c1a2444574b32dd7a283c53be248c5dbbf62e9