Numpy dtypes performance on squared euclidean distance

np_dtypes_sq_eucl.py

import time
import numpy as np
nl = np.random.randint(0, 1000, 9*10**6).reshape((3000,3000))

def sq_euclidean(X):
    XX = np.sum(X * X, axis=1)[:, np.newaxis]
    Y = X
    YY = XX.T
    distances = np.dot(X, Y.T)
    distances *= -2
    distances += XX
    distances += YY
    # remove floating point rounding errors
    np.maximum(distances, 0, distances)
    distances.flat[::distances.shape[0] + 1] = 0.0

for dtype in ['int16', 'int32', 'int64', 'float32', 'float64']:
    X = nl.astype(np.dtype(dtype), copy=True)
    t0 = time.time()
    _ = sq_euclidean(X)
    print(f'{dtype}: elapsed {time.time() - t0:.3f}s')