[tosa] failed to legalize operation 'arith.extsi'

[lundong]

What happened?

When we compile some mlir models via iree-compile, it returned the error messages like:

:0: error: loc("main"): 'arith.extsi' op operand #0 must be signless-fixed-width-integer-like, but got 'ui8'
:0: note: loc("main"): see current operation: %946 = "arith.extsi"(%arg423) : (ui8) -> i32

It has occured since iree-3.1.0rc20241206. Previous IREE revisions were okay.

Steps to reproduce your issue

1. Download the attached mobilenet_v1.mlir (which is converted from tflite model by iree-import-tflite)
2. iree-compile --iree-hal-target-backends=llvm-cpu path/to/mobilenet_v1.mlir
3. See error
   :0: error: loc("main"): 'arith.extsi' op operand #0 must be signless-fixed-width-integer-like, but got 'ui8'
   :0: note: loc("main"): see current operation: %946 = "arith.extsi"(%arg423) : (ui8) -> i32
   mobilenet_v1.zip

What component(s) does this issue relate to?

Compiler

Version information

iree-3.1.0rc20241206.

Additional context

This issue has occured since iree-3.1.0rc20241206. Previous IREE revisions were okay.

[IanWood1]

It appears that unsigned types (for example, ui8) are no longer legal for this op see llvm/llvm-project#118075. I'm not sure why there is no verifier to catch this, however.

[ScottTodd]

Oof. Mobilenet regression is not good. We're lacking test coverage for TOSA / TFLite (iree-org/iree-test-suites#5)

Commit range: iree-3.1.0rc20241205...iree-3.1.0rc20241206 . The PR that Ian linked is likely the culprit, coming in via 0f0cac5

[IanWood1]

@lundong where is this IR coming from? The linked IR contains the ui8 tosa.rescale ops.

[lundong]

It was converted from a Mobilenet V1 quantized TFLite model by iree-import-tflite. The TFLite model can be found at https://tfhub.dev/tensorflow/lite-model/mobilenet_v1_0.25_224_quantized/1/default/1?lite-format=tflite

[ScottTodd]

I don't think we've made any progress on fixing the regression, but I did just put together some tests that demonstrate the issue and will let us guard against similar regressions in the future: iree-org/iree-test-suites#59

[AGindinson]

It appears that unsigned types (for example, ui8) are no longer legal for this op

@IanWood1 I believe it's just a bug in the implementation that you've mentioned - the unsigned in/out types should remain valid at the TOSA dialect level. I'll create an upstream PR shortly.

[AGindinson]

Should be fixed by llvm/llvm-project#141096