Are we still using the ONNXToTOSA and ONNXToStablehlo flows in ONNX-MLIR?

[jagadeeshvx]

I have a question — are we still using the ONNXToTOSA and ONNXToStablehlo flows in ONNX-MLIR? I noticed that both flows have very limited operator support, so I’m wondering if they are still actively maintained or in use.

ONNXToTOSA - https://github.com/onnx/onnx-mlir/tree/main/src/Conversion/ONNXToTOSA
ONNXToStablehlo - https://github.com/onnx/onnx-mlir/tree/main/src/Conversion/ONNXToStablehlo

cc: @AlexandreEichenberger, @tungld

[jorickert]

At AMD, we are using the ONNXToTOSA flow. In our fork of onnx-mlir https://github.com/Xilinx/onnx-mlir we have extended operator support for this flow. We plan to upstream these changes, but need to discuss internally how and when best to do this.

[jagadeeshvx]

Thanks for the information, @jorickert. I have one more question—what is the difference between the ONNXToKrnl and ONNXToTOSA flows?

[AlexandreEichenberger]

ONNXToKrnl is one step to lower ONNX all the way to a binary implementing the model. After Krnl, we lower to affine/math/vector/scf/...
I believe the ONNXToTOSA flows takes ONNX to TOSA dialect. Then another compiler is used to lower TOSA to a lower executable representation.

[jagadeeshvx]

Thanks for the explanation @AlexandreEichenberger

[AlexandreEichenberger]

@jorickert let us know if there are blockers in the current main branch that makes that re-integration more difficult.