structured arrays slow to mask

I have 5 arrays of the same data type and the same length n (can be quite large). I need to regularly mask all of those array with the same mask, and I thought about what would be the fastest way. Fitting the data into one (n, 5) array works, and masking is equally fast with a (5, n) array.

I'd love to use a structured array so I can properly name the data in the rest of the code (e.g., arr["density"] instead of arr[:, 3]). Unfortunately, structured arrays mask way slower than the other two options:

Code to reproduce the plot:

import numpy
import perfplot
import random


def setup(n):
    a = numpy.random.rand(n, 5)
    b = numpy.ascontiguousarray(a.T)
    c = a.copy()
    c.dtype = [(f"col{k}", c.dtype) for k in range(a.shape[1])]

    idx = random.sample(range(0, n), 25)
    mask = numpy.ones(n, dtype=bool)
    mask[idx] = False
    return a, b, c, mask


def mask_rows(data):
    a, _, _, mask = data
    return a[mask]


def mask_named(data):
    _, _, c, mask = data
    return c[mask]


def mask_cols(data):
    _, b, _, mask = data
    return b[:, mask]


perfplot.show(
    setup=setup,
    kernels=[mask_rows, mask_cols, mask_named],
    n_range=[2 ** k for k in range(5, 20)],
    equality_check=None,
)

Start by running the provided Python reproduction and comparing boolean masking for the two-dimensional arrays and the structured array. Investigate the structured-array masking path responsible for the difference. Done means structured-array masking is no longer disproportionately slower, with the reproduction confirming the improvement.