rust-ndarray VS Rust-CUDA

I would love some examples of how to do non-trivial data interop between Rust and Python. My experience is that PyO3/Maturin is excellent when converting between simple datatypes but conversions get difficult when there are non-standard types, e.g. Python Numpy arrays or Rust ndarrays or whatever other custom thing.

Polars seems to have a good model where it uses the Arrow in memory format, which has implementations in Python and Rust, and makes a lot of the ndarray stuff easier. However, if the Rust libraries are not written with Arrow first, they become quite hard to work with. For example, there are many libraries written with https://github.com/rust-ndarray/ndarray, which is challenging to interop with Numpy.

(I am not an expert at all, please correct me if my characterizations are wrong!)

Sadly Ndarray does look a little abandoned to me: https://github.com/rust-ndarray/ndarray

The date of the last commit of [ndarray](https://github.com/rust-ndarray/ndarray) lies 6 month in the past while many recent issues are open and untouched.

Burn is different: it is built around the Backend trait which encapsulates tensor primitives. Even the reverse mode automatic differentiation is just a backend that wraps another one using the decorator pattern. The goal is to make it very easy to create optimized backends and support different devices and use cases. For now, there are only 3 backends: NdArray (https://github.com/rust-ndarray/ndarray) for a pure rust solution, Tch (https://github.com/LaurentMazare/tch-rs) for an easy access to CUDA and cuDNN optimized operations and the ADBackendDecorator making any backend differentiable. I am now refactoring the internal backend API to make it as easy as possible to plug in new ones.

Looks like it's an open request

https://github.com/rust-ndarray/ndarray/issues/281

Answer: you can’t with this crate. I implemented a dynamic n-dim solution myself but it uses views of integer indices that get copied to a new array, which have indexes to another flattened array in order to avoid duplication of possibly massive amounts of n-dimensional data; using the crate alone, copying all the array data would be unavoidable.

Ultimately I’ve had to make my own axis shifting and windowing mechanisms. But the crate is still a useful lib and continuing effort.

While I don’t mind getting into the weeds, these kinds of side efforts can really impact context focus so it’s just something to be aware of.

In principle you're trying to convert between columnar and row-based data layouts, something that happens fairly often in data science. I bet there's some hyper-efficient SIMD magic that could be invoked for these slicing operations (and maybe the iterator solution does exactly that). Might be worth taking a look at how the relevant Rust libraries like ndarray do it.

ndarray and numpy crates provide good way to operate on numpy ndarrays from python

I don't think any of the major ML projects have GPU acceleration because ndarray doesn't support it.

Not sure about ndarray: https://github.com/rust-ndarray/ndarray/issues/840

Source code [here](https://github.com/ihawn/RTracer) if anyone is interested in taking a look or giving feedback. As a side question, does anyone have any general advise on getting GPU compute working with rust? I tried [this project](https://github.com/Rust-GPU/Rust-CUDA) but had a bunch of issues (And it doesn't look like an active repo anyways)

Be warned, NON_BLOCKING streams do not fully synchronize with sync host to device copies. They are not guaranteed to actually finish by the time they return. Meaning its possible to initiate a copy, then initiate a kernel launch, and have the copy be unfinished by the time the kernel is launched. This caused so many confusing bugs that i personally decided to stop using NON_BLOCKING altogether in rust-cuda. https://github.com/Rust-GPU/Rust-CUDA/issues/15

- Cuda is not doing by FFI linking, instead is compiling CUDA code natively in Rust https://github.com/Rust-GPU/Rust-CUDA and even if it not complete as the C++ SDK is more than a toy

Given your criteria, you might want to consider (modern) C++.

* Fast - in many cases faster than Rust, although the difference is inconsequential relative to Python-to-Rust improvement I guess.

* _Really_ utilize CUDA, OpenCL, Vulcan etc. Specifically, Rust GPU is limited in its supported features, see: https://github.com/Rust-GPU/Rust-CUDA/blob/master/guide/src/... ...

* Host-side use of CUDA is at least as nice, and probably nicer, than what you'll get with Rust. That is, provided you use my own Modern C++ wrappers for the CUDA APIs: https://github.com/eyalroz/cuda-api-wrappers/ :-) ... sorry for the shameless self-plug.

* ... which brings me to another point: Richer offering of libraries for various needs than Rust, for you to possibly utilize.

* Easier to share than Rust. A target system is less likely to have an appropriate version of Rust and the surrounding ecosystem.

There are downsides, of course, but I was just applying your criteria.

Technically, yes. There are crates for OpenCL and CUDA, although official ROCm support does not exist yet.

Also you might not be aware but you can write cuda in rust now also https://github.com/Rust-GPU/Rust-CUDA

On the other hand CUDA is optimized to death and very good. Documentation and codes examples are everywhere (in C++ at least) but it is one more piece of software to install/configure and interact with from Rust. I don't know if Rust-CUDA is good or not. It's a WIP but the development seems stalled at this point (no commit since July)