This lets you parallelize a function designed for numpy inputs just by adding a few keyword arguments to apply_ufunc.

Example usage, to calculate rank correlation between two variables (this will probably turn into an example for the docs): ```python import numpy as np import xarray as xr import bottleneck

def covariance_gufunc(x, y): return ((x - x.mean(axis=-1, keepdims=True)) * (y - y.mean(axis=-1, keepdims=True))).mean(axis=-1)

def correlation_gufunc(x, y): return covariance_gufunc(x, y) / (x.std(axis=-1) * y.std(axis=-1))

def spearman_correlation_gufunc(x, y): x_ranks = bottleneck.rankdata(x, axis=-1) y_ranks = bottleneck.rankdata(y, axis=-1) return correlation_gufunc(x_ranks, y_ranks)

def spearman_correlation(x, y, dim): return xr.core.compuation.apply_ufunc( spearman_correlation_gufunc, x, y, input_core_dims=[[dim], [dim]], dask='parallelized', output_dtypes=[float]) We get a nice 5x speedup for this compute bound task: In [56]: rs = np.random.RandomState(0)

In [57]: array1 = xr.DataArray(rs.randn(1000, 100000), dims=['place', 'time']) # 800MB

In [58]: array2 = array1 + 0.5 * rs.randn(1000, 100000)

using one core, on numpy arrays

In [61]: %time _ = spearman_correlation(array1, array2, 'time') CPU times: user 21.6 s, sys: 2.84 s, total: 24.5 s Wall time: 24.9 s

using all my laptop's cores, with dask

In [63]: r = spearman_correlation(array1.chunk({'place': 10}), array2.chunk({'place': 10}), 'time')

In [64]: %time _ = r.compute() CPU times: user 30.9 s, sys: 1.74 s, total: 32.6 s Wall time: 4.59 s ```

Still needs examples in the documentation.

The next step is finally expose apply_ufunc as public API :).

• [x] Closes #585
• [x] Tests added / passed
• [x] Passes git diff upstream/master | flake8 --diff
• [x] Fully documented, including whats-new.rst for all changes and api.rst for new API

cc @mrocklin