Api -> Tosa operator mapping

[hemant456]

I’m looking to learn how each high level ML api from TF, Pytorch or Onnx is broken down using TOSA operators.

If such a mapping is public, could someone point me to mapping between high level api to TOSA primitives?

Secondly, I see TOSA operator list growing with releases. I’m curious what factored into extending primitive list?

[EricKunze]

Hi Hemant,

Welcome! The best place to look for breakdowns from the ML frameworks is in the MLIR compiler code which performs that task. You can find them in these locations:

• TensorFlow: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/compiler/mlir/tosa/transforms
  –mainly in legalize_tf.cc, legalize_common.cc, legalize_tfl.cc
• PyTorch has two paths
  – via the torch-mlir project: https://github.com/llvm/torch-mlir/tree/main/lib/Conversion/TorchToTosa
  – Executorch also supports a TOSA path: https://github.com/pytorch/executorch/tree/main/backends/arm
• ONNX has two options
  – onnx-mlir: https://github.com/onnx/onnx-mlir/tree/main/src/Conversion/ONNXToTOSA
  – The torch-mlir project can route ONNX through torch-mlir which can then be sent through the above listed torch-mlir path.

We try to keep the operator list stable, but you are right that there has been some growth. Growth usually comes as new network topologies come up and we find that TOSA is missing a building block. An example are SIN/COS, which we added to allow rotary position embeddings. We are aiming to release a 1.0 version soon, at which point we will guarantee backward compatibility, such that any 1.x TOSA release can continue to handle TOSA 1.0 networks.

Thanks,
Eric