`tosa.scatter` dimension requirement K >= W: Why do official test violates this rule?

MLIR
1

Hi everyone! Recently, I’m working with the tosa.scatter operation and noticed a potential inconsistency. I have read the official doc, and find the K>=W requirement. I have also checked the TOSA spec. The spec said K is the second dimension of values_in, and W is the second dimension of input. However, in the official test (llvm-project/mlir/test/Dialect/Tosa/ops.mlir at 72bb0a9a9c6fdde43e1e191f2dc0d5d2d46aff4e · llvm/llvm-project · GitHub, also pasted below), K=21, while W=26, which violates the requirement.

func.func @test_scatter(%arg0: tensor<13x21x3xf32>, %arg1: tensor<13x26xi32>, %arg2: tensor<13x26x3xf32>) -> tensor<13x21x3xf32> {
  %0 = tosa.scatter %arg0, %arg1, %arg2 : (tensor<13x21x3xf32>, tensor<13x26xi32>, tensor<13x26x3xf32>) -> tensor<13x21x3xf32>
  return %0 : tensor<13x21x3xf32>
}

I would like to know that is this expected? Or, am I misinterpreting anything?
Looking forward for your insight!

2

This looks like a bug in the test, and the validator doesn’t check for this, so it is not caught and reported. The test should get updated to not violate the spec requirements. We’re working on validation improvement, although I don’t know if this case specifically will get caught with the improvements.