Tabulate differences / progress on operators across StableHLO and its competitors

[SamuelMarks]

Back in the day this was basically my homepage: https://node.green

The specification in ML world isn't as agreed-upon as ECMAScript; with BLAS and Python-world NumPy and Scikit-learn having implementations everywhere from CUDA (cuBLAS, cuML) to Keras to PyTorch to TensorFlow's tf.experimental.numpy to JAX to ArrayFire to Dask &etc.

What I would like to see—and happy to assist in its original creation and scripts to maintain it—is a table with these columns:

Operation	NumPy	TensorFlow	JAX	⋱
dot (⋅)	✔	✔	✔	︙
⋱	…	…	…	⋰

maybe with extra detail like what version it was built, if it is partially implemented, if one impl has axis whilst the other has dimen, &etc.

It would be great if this was widely featured and placed on each framework's website. The advantage would be encouraging:

• compatibility
• interoperability
• portability
• good TODO lists for engineers that work on each project
• benchmarking different implementations
• understanding how to compare different implementations
  • e.g., one might be better on M [ARM-series] Mac's like MLX; whilst another may perform better on TPUs; whereas another still might perform better on—high-end or low-end—AMD or NVIDIA GPUs; &etc.
• developing tooling to automatically convert betwixt implementations… without having to rely on custom-written abstractions over abstractions adapter-pattern (DI inversion) frameworks like {Keras, ΦML, einops, …}.
  • Instead one should be able to take an arbitrary PyTorch codebase and convert it to a best-practice following JAX codebase without human intervention (or at least get >95% of the way to this)… and vice-vera.
  • The table could provide sufficient information to convert betwixt formats; and identify gaps in APIs so that a conversion isn't attempted but rather NotImplementedError is raised.

FWIW: There is other work in this direction like Python array API standard, your spec, ONNX, and others. But AFAIK no one is maintaining one of these tables. And maintaining one of these tables outside one of the big projects would be a recipe for obscurity and quickly becoming out-of-date.