An alternative implementation with nb.jitclass and array as field

[mishana]

A crisp mock example:

import numpy as np
import numba as nb
from numba.experimental import jitclass

spec = [
    ('array', nb.float64[:, :])
]

@jitclass(spec)
class A2D:
    def __init__(self, input_array):
        self.array = np.asarray(input_array, dtype=np.float64).reshape(-1, 2)
    
    @property
    def norm(self):
        return np.sqrt(self.array[:, 0]**2 + self.array[:, 1]**2)


>>> a_np = np.random.random(size=(1000, 2))
>>> a_2d = A2D(a_np)
>>> %timeit a2_d.norm
2.19 µs ± 40.2 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

>>> from vectorized2d import Array2D
>>> arr_2d = Array2D(a_np)
>>> %timeit arr_2d.norm
2.55 µs ± 44.1 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

[mishana]

I see two potential pros:

1. Possibly a simpler implementation.
2. Performance seems on-par and even better smh. Needs thorough testing after completing issue Performance Benchmark #25.

A concern: not extending np.array anymore. Is it better or worse??

[mishana]

All credit goes to @SharonLK for moisturizing my brain cells with the jitclass idea <3

[mishana]

I guess this feature isn't worth exploring, as long as inheritance between jitclasses is not supported: