[PYTHON] Extend NumPy with Rust

Benefits of Python

--Interactive data processing, analysis and visualization on Jupyter --Glue language usage --The Python interface is generally prepared by official and informal --Easy to build environment with pip / wheel or anaconda

Python problems

--Python itself is slow

Efforts to speed up

• Cython --Define a DSL that is close to Python that can be compiled, compile and execute
• Numba --Compile and run Python code in LLVM
• Theano --Build calculations with embedded DSLs on Python and run them on a fast backend
• FFI --Call an existing high-speed library (C / Fortran)

Python extensions and NumPy extensions

--Python C-API (CPython extension) - http://docs.python.jp/3/c-api/ --Define structures and functions that can be used in Python in C

• NumPy C-API
  • https://docs.scipy.org/doc/numpy/reference/c-api.html --In the first place, NumPy itself is implemented using Python C-API. --API for accessing structures added by NumPy

You need to use both of the two C-APIs

In what language do you want to extend NumPy?

• C/C++ --SWIG (NumPy C-API is used as it is)
• C++ --Boost.Python (including Boost.NumPy)
• Fortran
  • f2py
• Rust
  • rust-cpython / rust-numpy

Why Rust?

--Rust is a modern system language suitable for numerical calculations (I think) --Borrow Checker eliminates data races --C ++-like memory model-move semantics (no GC) --Typeclass generics --Powerful type inference - Native Thread --Optimized with LLVM -** I want to write in Rust **

rust-cpython

https://github.com/dgrunwald/rust-cpython

--Python C-API Rust wrapper --Manage reference counts --GIL abstraction --Launch Python interpreter --Publishing functions to Python

Call Python from Rust

extern crate cpython;

use cpython::{Python, PyDict, PyResult};

fn main() {
    let gil = Python::acquire_gil();
    hello(gil.python()).unwrap();
}

fn hello(py: Python) -> PyResult<()> {
    let sys = py.import("sys")?;
    let version: String = sys.get(py, "version")?.extract(py)?;

    let locals = PyDict::new(py);
    locals.set_item(py, "os", py.import("os")?)?;
    let user: String = py.eval("os.getenv('USER') or os.getenv('USERNAME')", None, Some(&locals))?.extract(py)?;

    println!("Hello {}, I'm Python {}", user, version);
    Ok(())
}

Calling Rust functions from Python

#[macro_use] extern crate cpython;

use cpython::{PyResult, Python};

// add bindings to the generated python module
// N.B: names: "librust2py" must be the name of the `.so` or `.pyd` file
py_module_initializer!(librust2py, initlibrust2py, PyInit_librust2py, |py, m| {
    try!(m.add(py, "__doc__", "This module is implemented in Rust."));
    try!(m.add(py, "sum_as_string", py_fn!(py, sum_as_string_py(a: i64, b:i64))));
    Ok(())
});

// logic implemented as a normal rust function
fn sum_as_string(a:i64, b:i64) -> String {
    format!("{}", a + b).to_string()
}

// rust-cpython aware function. All of our python interface could be
// declared in a separate module.
// Note that the py_fn!() macro automatically converts the arguments from
// Python objects to Rust values; and the Rust return value back into a Python object.
fn sum_as_string_py(_: Python, a:i64, b:i64) -> PyResult<String> {
    let out = sum_as_string(a, b);
    Ok(out)
}

rust-numpy

https://github.com/termoshtt/rust-numpy

--Made in GW (/ ・ ω ・) / --Published NumPy C-API to Rust based on rust-cpython --Introduce PyArray corresponding to numpy.ndarray --Convert to ʻArray` in rust-ndarray

rust-ndarray

https://github.com/bluss/rust-ndarray

--Linear algebra library in Rust --Stride method ndarray similar to NumPy

extern crate ndarray;
use ndarray::*;

// immutable example
fn axpy(a: f64, x: ArrayViewD<f64>, y: ArrayViewD<f64>) -> ArrayD<f64> {
    a * &x + &y
}

// mutable example (no return)
fn mult(a: f64, mut x: ArrayViewMutD<f64>) {
    x *= a;
}

rust-numpy

#[macro_use]
extern crate cpython;
extern crate numpy;

use numpy::*;
use cpython::{PyResult, Python, PyObject};

// wrapper of `axpy`
fn axpy_py(py: Python, a: f64, x: PyArray, y: PyArray) -> PyResult<PyArray> {
    let np = PyArrayModule::import(py)?;
    let x = x.as_array().into_pyresult(py, "x must be f64 array")?;
    let y = y.as_array().into_pyresult(py, "y must be f64 array")?;
    Ok(axpy(a, x, y).into_pyarray(py, &np))
}

// wrapper of `mult`
fn mult_py(py: Python, a: f64, x: PyArray) -> PyResult<PyObject> {
    let x = x.as_array_mut().into_pyresult(py, "x must be f64 array")?;
    mult(a, x);
    Ok(py.None()) // Python function must returns
}

/* Define module "_rust_ext" */
py_module_initializer!(_rust_ext, init_rust_ext, PyInit__rust_ext, |py, m| {
    m.add(py, "__doc__", "Rust extension for NumPy")?;
    m.add(py, "axpy", py_fn!(py, axpy_py(a: f64, x: PyArray, y: PyArray)))?;
    m.add(py, "mult", py_fn!(py, mult_py(a: f64, x: PyArray)))?;
    Ok(())
});

setuptools-rust

--rust-A library for handling cpython extensions with setuptools

setup.py
extensions/Cargo.toml
           src/lib.rs
rust_ext/__init__.py

setup.py

from setuptools import setup
from setuptools_rust import RustExtension

setup(name='rust_ext',
      version='1.0',
      rust_extensions=[
          RustExtension('rust_ext._rust_ext', 'extensions/Cargo.toml')],
      packages=['rust_ext'],
      zip_safe=False)

rust_ext/__init__.py

from ._rust_ext import *

Call rust-numpy

python setup.py install

import rust_ext
import numpy as np

x = np.array([1.0, 2.0])
y = np.array([2.0, 3.0])
rust_ext.axpy(3, x, y)

(*'▽') It worked!

Finally: When Rust is painful

• rust-jp.slack.com
• ja.stackoverflow.com/questions/tagged/rust --Kobe Rust Association