`normalize_total` with numba

[Intron7]

What kind of feature would you like to request?

Other?

Please describe your wishes

This would speed up normalization

[ilan-gold]

Waiting for potential work already done on this by intel folks (@ashish615 have you guys worked on this?)

[ashish615]

@Intron7 code for below link may work for only csr matrix.
https://github.com/IntelLabs/Open-Omics-Acceleration-Framework/blob/main/pipelines/single-cell-RNA-seq-analysis/notebooks/fastpp.py#L499-L522

[ilan-gold]

Implement with or without Intel PR

[selmanozleyen]

So the implementation from intel would replace the axis_mul_or_truediv function we use as last step here:

scanpy/src/scanpy/preprocessing/_normalization.py

Lines 28 to 52 in 834159a

	def _normalize_data(X, counts, after=None, *, copy: bool = False):
	X = X.copy() if copy else X
	if issubclass(X.dtype.type, int | np.integer):
	X = X.astype(np.float32) # TODO: Check if float64 should be used
	if after is None:
	if isinstance(counts, DaskArray):
	def nonzero_median(x):
	return np.ma.median(np.ma.masked_array(x, x == 0)).item()
	after = da.from_delayed(
	dask.delayed(nonzero_median)(counts),
	shape=(),
	meta=counts._meta,
	dtype=counts.dtype,
	)
	else:
	counts_greater_than_zero = counts[counts > 0]
	after = np.median(counts_greater_than_zero, axis=0)
	counts = counts / after
	out = X if isinstance(X, np.ndarray | CSBase) else None
	return axis_mul_or_truediv(
	X, counts, op=truediv, out=out, allow_divide_by_zero=False, axis=0
	)

Because everything done until that point is to find median in case there is no fixed target sum given. So I would essentially have to utilize numba in axis_mul_or_truediv, does that sound right @ilan-gold @flying-sheep ? If so I would proceed on that

[ilan-gold]

@selmanozleyen That does sound correct, although I have not looked into it closely before just now, and there is not much of a bread crumb trail to pick at since the Intel code is so different. So I don't want to say "yes for sure" but it does seem that way. It's possible axis_sum could be another place to use numba, I'm really not sure. BTW a good place for something like "fast axis multiplication" or "axis sum" could be https://github.com/scverse/fast-array-utils (and in fact maybe should be there). But prototyping here and getting the actual numba function ready would make sense. Maybe @Intron7 wants to add something as to which part of normalize_total would be "numba-able" if not those mentioned?