[Mlir-commits] [mlir] [mlir][Linalg] use linalg.reduce to simplify the mergeReductions in partialReductionInterface (PR #94579)

[zhicong zhong]

zhczhong wrote:

> I just have a question from downstream use of this. If I run generalization of the `linalg.reduce` op do we get back the same `linalg.generic` generated?

Yes, the linalg generalization can convert the `linalg.reduce` to `linalg.generic` in the same form as the original implementation.
```mlir
func.func @test(%input: tensor<16x32x64xf32>,
                  %init: tensor<16x64xf32>) -> tensor<16x64xf32> {
  %reduce = linalg.reduce
      ins(%input:tensor<16x32x64xf32>)
      outs(%init:tensor<16x64xf32>)
      dimensions = [1]
      (%in: f32, %out: f32) {
        %0 = arith.addf %out, %in: f32
        linalg.yield %0: f32
      }
  func.return %reduce : tensor<16x64xf32>
}
module attributes {transform.with_named_sequence} {
  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
    %0 = transform.structured.match interface{LinalgOp} in %arg1 : (!transform.any_op) -> !transform.any_op
    %1 = transform.structured.generalize %0 : (!transform.any_op) -> !transform.any_op
    transform.yield
  }
}
```

will be converted to

```mlir
#map1 = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
#map2 = affine_map<(d0, d1, d2) -> (d0, d2)>
func.func @test(%arg0: tensor<16x32x64xf32>, %arg1: tensor<16x64xf32>) -> tensor<16x64xf32> {
  %0 = linalg.generic {indexing_maps = [#map1, #map2], iterator_types = ["parallel", "reduction", "parallel"]} ins(%arg0 : tensor<16x32x64xfarg0 : tensor<16x32x64xf32>) outs(%arg1 : tensor<16x64xf32>) {                                                       
  ^bb0(%in: f32, %out: f32):
    %1 = arith.addf %out, %in : f32
    linalg.yield %1 : f32
  } -> tensor<16x64xf32>
  return %0 : tensor<16x64xf32>
}
```

https://github.com/llvm/llvm-project/pull/94579