I would be seriously grateful for a Swift <--> C++ integration friend

[dougdew64]

I am attempting to implement a swift equivalent of the c++ tutorial application.

I am getting my butt kicked in Xcode trying to figure out how to correctly structure and configure my projects in order to make Xcode happy. After several hours, I found a structure and configuration which Xcode doesn't complain about.

Now I'm encountering a surprising compilation failure on my first line of tutorial code. The swift compiler complains that there's no mlx.core.array (mlx::core::array) available. I've attempted this compilation from the command line and experienced the same failure.

Interestingly, the Xcode intellisense seems to be aware of mlx.core.array.

[RahulBhalley]

This would be really interesting as S4TF was killed unexpectedly.

• Author of Deep Learning with Swift for TensorFlow book here

[dougdew64]

@RahulBhalley you might find this discussion interesting: #15

[RahulBhalley]

Hey! Any success in running MLX with C++ project in Xcode?

[dougdew64]

@RahulBhalley I used MLX, C++ and Xcode a while back. #86

[dougdew64]

I've been pretty much stuck since then. I had been intending to add higher-level C++ and eventually implement an all-C++ version of the LLM example. However, a few weeks ago @awni and @davidkoski suggested that the Swift API would be available soon. So, I paused my effort to await the Swift API. When the Swift API is available, then I'll resume my effort while using the Swift API as a guide for my C++ effort. For example, I want to see how details such as array slicing are implemented in Swift before attempting the same in C++.

[davidkoski]

Depending on what you mean, you can take a look at the implementation of the python binding:

• https://github.com/ml-explore/mlx/blob/main/python/src/indexing.cpp#L378

mlx_get_item_slice is array[1:10] in python. This in turn calls slice() in the C++ back end:

• https://github.com/ml-explore/mlx/blob/main/mlx/ops.cpp#L439

The swift implementation could call the same code in the C++ back end for its implementation of array[1 ..< 10]

That doesn't really answer how to call that from swift (larger issue) but it would let you write a C++ version of python code that was slicing (which is what I understood the question to be asking).

[dougdew64]

Thanks @davidkoski.

I had been thinking along the lines which you described, but wanted to first see how you solve the same problem(s) in the Swift binding.

Are you implementing in Swift the same kind of logic which is contained in indexing.cpp?

[davidkoski]

Some of it -- the python API allows you to mix all of the different indexing operations at once. The swift version is typed and you would just compose them or use different operators. If you look at mlx_get_item_nd() that is what the python side does internally.

In practice, e.g. porting some python code for testing, this has worked out ok. It doesn't preclude doing subscript(Any...) in the future. I think in C++ it might be a little easier with std::variant.

[dougdew64]

Thanks @davidkoski