Clear up broadcasting

Author: dcherian

xarray/core/missing.py

@@ -3,6 +3,7 @@
 import datetime as dt
 import itertools
 import warnings
+from collections import ChainMap
 from collections.abc import Callable, Generator, Hashable, Sequence
 from functools import partial
 from numbers import Number
@@ -710,59 +711,66 @@ def interpolate_variable(
         func, kwargs = _get_interpolator_nd(method, **kwargs)
 
     in_coords, result_coords = zip(*(v for v in indexes_coords.values()), strict=True)
-    # broadcast out manually to minize confusing behaviour
-    broadcast_result_coords = broadcast_variables(*result_coords)
-    result_dims = broadcast_result_coords[0].dims
 
     # input coordinates along which we are interpolation are core dimensions
     # the corresponding output coordinates may or may not have the same name,
     # so `all_in_core_dims` is also `exclude_dims`
     all_in_core_dims = set(indexes_coords)
 
+    result_dims = OrderedSet(itertools.chain(*(_.dims for _ in result_coords)))
+    result_sizes = ChainMap(*(_.sizes for _ in result_coords))
+
     # any dimensions on the output that are present on the input, but are not being
-    # interpolated along are broadcast or loop dimensions along which we automatically
-    # vectorize. Consider the problem in
-    # https://github.com/pydata/xarray/issues/6799#issuecomment-2474126217
+    # interpolated along are dimensions along which we automatically vectorize.
+    # Consider the problem in https://github.com/pydata/xarray/issues/6799#issuecomment-2474126217
     # In the following, dimension names are listed out in [].
     # # da[time, q, lat, lon].interp(q=bar[lat,lon]). Here `lat`, `lon`
-    # are input dimensions, present on the output, along which we vectorize.
-    # We track these as "result broadcast dimensions".
+    # are input dimensions, present on the output, but are not the coordinates
+    # we are explicitly interpolating. These are the dimensions along which we vectorize.
     # `q` is the only input core dimensions, and changes size (disappears)
     # so it is in exclude_dims.
-    result_broadcast_dims = set(
-        itertools.chain(dim for dim in result_dims if dim not in all_in_core_dims)
-    )
+    vectorize_dims = (result_dims - all_in_core_dims) & set(var.dims)
 
     # remove any output broadcast dimensions from the list of core dimensions
-    output_core_dims = tuple(d for d in result_dims if d not in result_broadcast_dims)
+    output_core_dims = tuple(d for d in result_dims if d not in vectorize_dims)
     input_core_dims = (
         # all coordinates on the input that we interpolate along
         [tuple(indexes_coords)]
         # the input coordinates are always 1D at the moment, so we just need to list out their names
         + [tuple(_.dims) for _ in in_coords]
         # The last set of inputs are the coordinates we are interpolating to.
-        # These have been broadcast already for ease.
-        + [output_core_dims] * len(result_coords)
+        + [
+            tuple(d for d in coord.dims if d not in vectorize_dims)
+            for coord in result_coords
+        ]
     )
-    output_sizes = {k: broadcast_result_coords[0].sizes[k] for k in output_core_dims}
+    output_sizes = {k: result_sizes[k] for k in output_core_dims}
 
     # scipy.interpolate.interp1d always forces to float.
     dtype = float if not issubclass(var.dtype.type, np.inexact) else var.dtype
     result = apply_ufunc(
         _interpnd,
         var,
         *in_coords,
-        *broadcast_result_coords,
+        *result_coords,
         input_core_dims=input_core_dims,
         output_core_dims=[output_core_dims],
         exclude_dims=all_in_core_dims,
         dask="parallelized",
-        kwargs=dict(interp_func=func, interp_kwargs=kwargs),
+        kwargs=dict(
+            interp_func=func,
+            interp_kwargs=kwargs,
+            # we leave broadcasting up to dask if possible
+            # but we need broadcasted values in _interpnd, so propagate that
+            # context (dimension names), and broadcast there
+            # This would be unnecessary if we could tell apply_ufunc
+            # to insert size-1 broadcast dimensions
+            result_coord_core_dims=input_core_dims[-len(result_coords) :],
+        ),
         # TODO: deprecate and have the user rechunk themselves
         dask_gufunc_kwargs=dict(output_sizes=output_sizes, allow_rechunk=True),
         output_dtypes=[dtype],
-        # if there are any broadcast dims on the result, we must vectorize on them
-        vectorize=bool(result_broadcast_dims),
+        vectorize=bool(vectorize_dims),
         keep_attrs=True,
     )
     return result
@@ -787,7 +795,11 @@ def _interp1d(
 
 
 def _interpnd(
-    data: np.ndarray, *coords: np.ndarray, interp_func: InterpCallable, interp_kwargs
+    data: np.ndarray,
+    *coords: np.ndarray,
+    interp_func: InterpCallable,
+    interp_kwargs,
+    result_coord_core_dims,
 ) -> np.ndarray:
     """
     Core nD array interpolation routine.
@@ -801,10 +813,12 @@ def _interpnd(
     # Convert everything to Variables, since that makes applying
     # `_localize` and `_floatize_x` much easier
     x = [Variable([f"dim_{nconst + dim}"], _x) for dim, _x in enumerate(coords[:n_x])]
-    new_x = [
-        Variable([f"dim_{ndim + dim}" for dim in range(_x.ndim)], _x)
-        for _x in coords[n_x:]
-    ]
+    new_x = broadcast_variables(
+        *(
+            Variable(dims, _x)
+            for dims, _x in zip(result_coord_core_dims, coords[n_x:], strict=True)
+        )
+    )
     var = Variable([f"dim_{dim}" for dim in range(ndim)], data)
 
     if interp_kwargs.get("method") in ["linear", "nearest"]: