[all-commits] [llvm/llvm-project] be58d4: [mlir] Canonicalize dynamic tensor.pad ops with co...

[gpetters94 via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: be58d484cb315028049255d9bb29ce2c7bffc983
      https://github.com/llvm/llvm-project/commit/be58d484cb315028049255d9bb29ce2c7bffc983
  Author: George Petterson <gpetters at protonmail.com>
  Date:   2023-02-03 (Fri, 03 Feb 2023)

  Changed paths:
    M mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
    M mlir/test/Dialect/Tensor/canonicalize.mlir

  Log Message:
  -----------
  [mlir] Canonicalize dynamic tensor.pad ops with constant inputs

This commit adds a canonicalization pattern for tensor.pad which changes the output type to static at each dimension where the input shape is static and the high and low operands are constants. This corrects an issue arising in Torch-MLIR where pad ops would sometimes introduce dynamic shapes unnecessarily.

Reviewed By: raikonenfnu

Differential Revision: https://reviews.llvm.org/D143135