Optimize padding for coarsening.

asv_bench/benchmarks/coarsen.py

@@ -0,0 +1,47 @@
+import numpy as np
+
+import xarray as xr
+
+from . import randn
+
+# Sizes chosen to test padding optimization
+nx_padded = 4003  # Not divisible by 10 - requires padding
+ny_padded = 4007  # Not divisible by 10 - requires padding
+
+nx_exact = 4000  # Divisible by 10 - no padding needed
+ny_exact = 4000  # Divisible by 10 - no padding needed
+
+window = 10
+
+
+class Coarsen:
+    def setup(self, *args, **kwargs):
+        # Case 1: Requires padding on both dimensions
+        self.da_padded = xr.DataArray(
+            randn((nx_padded, ny_padded)),
+            dims=("x", "y"),
+            coords={"x": np.arange(nx_padded), "y": np.arange(ny_padded)},
+        )
+
+        # Case 2: No padding required
+        self.da_exact = xr.DataArray(
+            randn((nx_exact, ny_exact)),
+            dims=("x", "y"),
+            coords={"x": np.arange(nx_exact), "y": np.arange(ny_exact)},
+        )
+
+    def time_coarsen_with_padding(self):
+        """Coarsen 2D array where both dimensions require padding."""
+        self.da_padded.coarsen(x=window, y=window, boundary="pad").mean()
+
+    def time_coarsen_no_padding(self):
+        """Coarsen 2D array where dimensions are exact multiples (no padding)."""
+        self.da_exact.coarsen(x=window, y=window, boundary="pad").mean()
+
+    def peakmem_coarsen_with_padding(self):
+        """Peak memory for coarsening with padding on both dimensions."""
+        self.da_padded.coarsen(x=window, y=window, boundary="pad").mean()
+
+    def peakmem_coarsen_no_padding(self):
+        """Peak memory for coarsening without padding."""
+        self.da_exact.coarsen(x=window, y=window, boundary="pad").mean()

doc/whats-new.rst

@@ -51,6 +51,8 @@ Performance
 - Speedup and reduce memory usage of :py:func:`concat`. Magnitude of improvement scales
   with size of the concatenation dimension. By `Deepak Cherian <https://github.com/dcherian>`_.
   :issue:`10864` :pull:`10866`.
+- Speedup and reduce memory usage when coarsening along multiple dimensions.
+  By `Deepak Cherian <https://github.com/dcherian>`_. :pull:`10921`.
 
 Documentation
 ~~~~~~~~~~~~~

xarray/core/variable.py

@@ -2280,6 +2280,7 @@ def coarsen_reshape(self, windows, boundary, side):
                 )
 
         variable = self
+        pad_widths = {}
         for d, window in windows.items():
             # trim or pad the object
             size = variable.shape[self._get_axis_num(d)]
@@ -2300,16 +2301,16 @@ def coarsen_reshape(self, windows, boundary, side):
                 pad = window * n - size
                 if pad < 0:
                     pad += window
-                if side[d] == "left":
-                    pad_width = {d: (0, pad)}
-                else:
-                    pad_width = {d: (pad, 0)}
-                variable = variable.pad(pad_width, mode="constant")
+                elif pad == 0:
+                    continue
+                pad_widths[d] = (0, pad) if side[d] == "left" else (pad, 0)
             else:
                 raise TypeError(
                     f"{boundary[d]} is invalid for boundary. Valid option is 'exact', "
                     "'trim' and 'pad'"
                 )
+        if pad_widths:
+            variable = variable.pad(pad_widths, mode="constant")
 
         shape = []
         axes = []