Unary broadcast lowering

[KavithaTipturMadhu]

No description provided.

[rengolin]

You have huge gaps in your testing. You need to make sure the tests are correct (I'm not so sure about that), that they follow the path you expect (linalg-to-vector and vector-to-xsmm), and that the output is correct.

You did some of those things for different integration tests. You need to do all of those things for all integration tests. You also need negative conversion tests (dims that don't match, reducing rank, etc), and a more comprehensive functionality testing (more patterns like 1D to 2D, 2D to 3D, 1D to 3D, etc).

It would also be interesting to have a more elaborate test, where you start adding all of the supported patterns. For now, one that has a transpose followed by a broadcast and another that is the other way round. You should match both patterns on both cases, and create two function calls on both.

[rengolin]

Is this really a 2D column broadcast?

[KavithaTipturMadhu]

It is a column broadcast, which is what the function is called, the file should probably be renamed broadcast with 2D destination or something, can you please suggest an appropriate name for the test file?

[adam-smnk]

+1 on more test coverage or rather more specific matchers
Our tests right now surely don't cover all vector.broadcast semantics.

As an example:

func.func @entry(%arg0: memref<4x1xf32>, %arg1: memref<4x2xf32>) -> memref<4x2xf32> {
  %c0 = arith.constant 0 : index
  %cst = arith.constant 0.000000e+00 : f32
  %0 = vector.transfer_read %arg0[%c0, %c0], %cst {in_bounds = [true, true]} : memref<4x1xf32>, vector<4x1xf32>
  %1 = vector.broadcast %0 : vector<4x1xf32> to vector<4x2xf32>
  vector.transfer_write %1, %arg1[%c0, %c0] {in_bounds = [true, true]}
    : vector<4x2xf32>, memref<4x2xf32>
  return %arg1 : memref<4x2xf32>
}

This broadcast fails at runtime xsmm JIT step. I'm sure we don't need this variant currently but it'd be best not to open ourselves to new pitfalls.

[adam-smnk]

As a side note, it could be that our current lowering from linalg is also incomplete. I don't want to feature/scope creep this PR.
Let's just agree on some sufficient state and possibly note down future necessary improvements.