mlir-opt --tosa-to-linalg-pipeline fails in validation: element type 'f32' is not legal

[PerMildner]

It seems mlir-opt --tosa-to-linalg-pipeline unconditionally uses the Basic Inference profile, and therefore only allows integer types.

The help-text for --tosa-to-linalg-pipeline says

The default pipeline for converting TOSA operators to the equivalent operations using the tensor operations in LinAlg as well as LinAlg named operations.

so I would have assumed it should be usable also for floating point tensors.

I could not find any tests using --tosa-to-linalg-pipeline except tests that are supposed to fail. Other tests seems build a stripped down version of --tosa-to-linalg-pipeline using explicit --pass-pipeline.

The symptom is:

$ cat ../bad_tosa.mlir
module attributes {llvm.data_layout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128", llvm.target_triple = "x86_64-unknown-linux-gnu"} {
  func.func @main_graph() -> (tensor<1xf32>) {
    %0 = "tosa.const"() <{value = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32>
    return %0 : tensor<1xf32>
  }
}

$ build/bin/mlir-opt --tosa-to-linalg-pipeline ../bad_tosa.mlir
../bad_tosa.mlir:3:10: error: 'tosa.const' op is not profile-aligned: element type 'f32' is not legal
    %0 = "tosa.const"() <{value = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32>
         ^
../bad_tosa.mlir:3:10: note: see current operation: %0 = "tosa.const"() <{value = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32>
../bad_tosa.mlir:4:5: error: 'func.return' op is not profile-aligned: element type 'f32' is not legal
    return %0 : tensor<1xf32>
    ^
../bad_tosa.mlir:4:5: note: see current operation: "func.return"(%0) : (tensor<1xf32>) -> ()

Commit 2bcbe40.

I have attached a full transcript.

bad_tosa.txt