issue_comments

Over in https://github.com/pydata/xarray/issues/7344#issuecomment-1336299057 @shoyer

That said -- we could also switch to smarter NumPy based algorithms to implement most moving window calculations, e.g,. using np.nancumsum for moving window means.

After some digging, this would involve using "summed area tables" which have been generalized to nD, and can be used to compute all our built-in reductions (except median). Basically we'd store the summed area table (repeated np.cumsum) and then calculate reductions using binary ops (mostly subtraction) on those tables.

This would be an intermediate level project but we could implement it incrementally (start with sum for example). One downside is the potential for floating point inaccuracies because we're taking differences of potentially large numbers.

cc @aulemahal

I think I may have found a way to make the variance/standard deviation calculation more memory efficient, but I don't know enough about writing the sort of code that would be needed for a PR.

I basically wrote out the calculation for variance trying to only use the functions that have already been optimsed. Derived from:

$$ var = \frac{1}{n} \sum_{i=1}^{n} (x_i - \mu)^2 $$

$$ var = \frac{1}{n} \left( (x_1 - \mu)^2 + (x_2 - \mu)^2 + (x_3 - \mu)^2 + ... \right) $$

$$ var = \frac{1}{n} \left(x_1^2 -2x_1\mu + \mu^2 + \ x_2^2 -2x_2\mu + \mu^2 + \ x_3^2 -2x_3\mu + \mu^2 + ... \right) $$

$$ var = \frac{1}{n} \left( \sum_{i=1}^{n} x_i^2 - 2\mu\sum_{i=1}^{n} x_i + n\mu^2 \right)$$

I coded this up and demonstrate that it uses approximately 10% of the memory as the current .var() implementation:

```python %load_ext memory_profiler

import numpy as np import xarray as xr

temp = xr.DataArray(np.random.randint(0, 10, (5000, 500)), dims=("x", "y"))

def new_var(da, x=10, y=20): # Defining the re-used parts roll = da.rolling(x=x, y=y) mean = roll.mean() count = roll.count() # First term: sum of squared values term1 = (da2).rolling(x=x, y=y).sum() # Second term cross term sum term2 = -2 * mean * roll.sum() # Third term 'sum' of squared means term3 = count * mean2 # Combining into the variance var = (term1 + term2 + term3) / count return var

def old_var(da, x=10, y=20): roll = da.rolling(x=x, y=y) var = roll.var() return var

%memit new_var(temp) %memit old_var(temp) ```

peak memory: 429.77 MiB, increment: 134.92 MiB peak memory: 5064.07 MiB, increment: 4768.45 MiB

I wanted to double check that the calculation was working correctly:

python print((var_o.where(~np.isnan(var_o), 0) == var_n.where(~np.isnan(var_n), 0)).all().values) print(np.allclose(var_o, var_n, equal_nan = True))

False True

I think the difference here is just due to floating point errors, but maybe someone who knows how to check that in more detail could have a look.

The standard deviation can be trivially implemented from this if the approach works.

This is already done for counts, correct? Here:

https://github.com/pydata/xarray/blob/1597e3a91eaf96626725987d23bbda2a80d2bae7/xarray/core/rolling.py#L370-L382

This should work for most of the reductions (and is a bit similar to what is done in weighted for mean and sum):

• [x] count: isnull() -> rolling -> sum
• [x] argmax: fillna(-inf) -> rolling -> argmax
• [x] argmin: fillna(inf) -> rolling -> argmin
• [x] max: fillna(-inf) -> rolling -> max (not sure about this one, need to be careful with the dtype)
• [x] min: fillna(inf) -> rolling -> min (dito)
• [x] mean: fillna(0) -> rolling -> sum / count (ensure nan if count == 0)
• [x] prod: fillna(1) -> rolling -> prod
• [x] sum: fillna(0) -> rolling -> sum
• [ ] var: fillna(0) -> rolling -> possible (?) but a bit more involved
• [ ] std: sqrt(var)
• [ ] median: probably not possible

I think this should not be too difficult, the thing is that rolling itself is already quite complicated

I just saw that numpy 1.20 introduces stride_tricks.sliding_window_view. I have not looked at this yet. Just leaving this here for reference.

https://numpy.org/devdocs/reference/generated/numpy.lib.stride_tricks.sliding_window_view.html#numpy.lib.stride_tricks.sliding_window_view

https://numpy.org/devdocs/release/1.20.0-notes.html#sliding-window-view-provides-a-sliding-window-view-for-numpy-arrays

https://github.com/numpy/numpy/pull/17394

@mathause Oh, I missed this issue. Yes, this is implemented only for count.

the thing is that rolling itself is already quite complicated

Agreed. We need to clean this up.

One possible option would be to drop support of bottleneck. This does not work for nd-rolling and if we implement the nd-nanreduce, the speed should be comparable with bottleneck.