Speed up notch filter spectrum fit

[cbrnr]

Adding @cache to _get_window_thresh results in a speedup of ~1.85x on my machine (15.48s vs. 8.37s) in the following benchmark:

from pathlib import Path
from time import perf_counter

import numpy as np

import mne
from mne.datasets import sample

mne.set_log_level("WARNING")

fname = Path(sample.data_path()) / "MEG" / "sample" / "sample_audvis_raw.fif"

raw = mne.io.read_raw_fif(fname, preload=True)
raw.crop(tmin=0, tmax=60)

start = perf_counter()
raw.notch_filter(freqs=np.arange(50, 251, 50), method="spectrum_fit")
elapsed = perf_counter() - start
print(f"{elapsed=:.3f}s")

[larsoner]

Two small potential ideas for the future work, but not blockers I think

[larsoner]

At some point we might want to make this conditional on the size. If you have like a 2GB file this is going to be painful, and doing it in chunks would be much less so (and not much slower).

[cbrnr]

I'll revert this then - it turns out for the notch filter use case, it always receives a single channel anyway, but this could affect other call sites which I didn't think of.

[larsoner]

Maybe better here would be a lru_cache with a maxsize (100?) to prevent it growing without bound. But here the risk is probably low in practice so okay to stay as is (maybe fix next time someone is working in the file)

[cbrnr]

You are right, I'll do that in a follow-up PR. The risk is minimal in practice, but better to be safe. I'll just use the default of maxsize=128.

[larsoner]

Thanks @cbrnr !