Apple Silicon Support (ORT and CoreML)

[danielholanda]

Description

This PR adds Apple Silicon support to ORT and creates a new runtime for CoreML.

PR was tested on MacOS v12.7.1 (21G920) on a MacBook M1.

Setup

I only tested this on my local machine, so I'm curious to understand how smooth the installation process is.

I got an OSError with Lightgbm, which was solved by running conda install lightgbm. Not sure if others will encounter the same.

Testing

cd into the timm folder and run turnkey resnet18.py --device apple_silicon --runtime coreml

To do before merging this PR

• Set CoreML as the default runtime of the apple_silicon device
• Add disclaimer on best performance not being guaranteed (applies to all runtimes)

Future Work (Issues to be created)

• Investigate if using Swift as the coreML runtime leads to higher performance
• Add bash size as a runtime argument (also applies to ORT)
• Use different inputs when benchmarking instead of the same input (also applies to ORT)
• Add tests: Blocked by lack of available Apple Silicon hardware in the cloud

[danielholanda]

This naming convention was modified as it was too restrictive. We certainly don't want all models going to onnx/model.onnx if the output of the build process is not an onnx model.

Please note that this is not a breaking change according to our plugin contract, as the signature of BaseRT has not changed (apart from optional args added).