[all-commits] [llvm/llvm-project] de5cef: [mlir][vector] Extend `combineContractAndBroadcast...

[Andrzej Warzyński via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: de5cef3127047836fceb416f8083942783834489
      https://github.com/llvm/llvm-project/commit/de5cef3127047836fceb416f8083942783834489
  Author: Andrzej Warzyński <andrzej.warzynski at arm.com>
  Date:   2026-07-17 (Fri, 17 Jul 2026)

  Changed paths:
    M mlir/include/mlir/Dialect/Vector/IR/VectorOps.h
    M mlir/include/mlir/Dialect/Vector/IR/VectorOps.td
    M mlir/lib/Dialect/Vector/IR/VectorOps.cpp
    M mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
    M mlir/test/Dialect/Vector/vector-reduce-to-contract.mlir

  Log Message:
  -----------
  [mlir][vector] Extend `combineContractAndBroadcast` to accept `vector.shape_cast` (#208752)

`combineContractAndBroadcast` is a pattern that matches
`vector.contract` operations whose inputs are defined by
`vector.broadcast`, e.g.:

```mlir
%0 = vector.broadcast %lhs : vector<8x4xi32> to vector<1x8x4xi32>
%1 = vector.broadcast %rhs : vector<8x4xi32> to vector<1x8x4xi32>
%result = vector.contract {
    indexing_maps = [#map0, #map1, #map2],
    iterator_types = ["reduction", "parallel", "parallel", "reduction"],
    kind = #vector.kind<add>
  } %0, %1, %acc : vector<1x8x4xi32>, vector<1x8x4xi32> into vector<8x8xi32>
```

The pattern folds away the reduction over a unit dimension introduced by
the defining operations, producing a simplified `vector.contract`:

```mlir
vector.contract {
  indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP2]]]
  iterator_types = ["parallel", "parallel", "reduction"]
} %0, %1, %acc : vector<8x4xi32>, vector<8x4xi32> into vector<8x8xi32>
```

This change extends the pattern to also accept broadcast-like
`vector.shape_cast` operations:

```mlir
%0 = vector.shape_cast %lhs : vector<8x4xi32> to vector<1x8x4xi32>
%1 = vector.shape_cast %rhs : vector<8x4xi32> to vector<1x8x4xi32>
%result = vector.contract {
    indexing_maps = [#map0, #map1, #map2],
    iterator_types = ["reduction", "parallel", "parallel", "reduction"],
    kind = #vector.kind<add>
  } %0, %1, %acc : vector<1x8x4xi32>, vector<1x8x4xi32> into vector<8x8xi32>
```

Note that `mlir::vector::isBroadcastableTo` was only moved to make it
available to the newly introduced `ShapeCastOp` helper,
`ShapeCastOp::isBroadcastLike()`.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications