Address review: parametrize dask+cupy descending test, note filter overshoot (#3627)

Author: brendancol

xrspatial/kde.py

@@ -391,6 +391,8 @@ def _filter_points_to_tile(xs, ys, ws, tile_x0, tile_y0, dx, dy,
     # dx/dy may be negative (descending coordinates), so order the tile
     # edges with min/max before widening by the cutoff.  The unordered
     # version inverted the interval and dropped the points (#3627).
+    # tile_x1/tile_y1 overshoot the last pixel centre by one spacing,
+    # which keeps the filter conservative (never drops a contributor).
     tile_x1 = tile_x0 + tile_cols * dx
     tile_y1 = tile_y0 + tile_rows * dy
     x_lo = min(tile_x0, tile_x1) - cutoff

xrspatial/tests/test_kde.py

@@ -534,13 +534,17 @@ def test_cupy_matches_numpy(self, point_cluster, simple_grid, kernel):
 class TestDaskCupyDescending:
     """dask+cupy must not drop points on descending templates (#3627)."""
 
-    def test_dask_cupy_matches_numpy_descending_y(self, point_cluster):
+    @pytest.mark.parametrize('desc_y,desc_x', [(True, False), (False, True)])
+    def test_dask_cupy_matches_numpy_descending(self, point_cluster,
+                                                desc_y, desc_x):
         import cupy
         x, y = point_cluster
+        ys = np.linspace(4, -4, 16) if desc_y else np.linspace(-4, 4, 16)
+        xs = np.linspace(4, -4, 16) if desc_x else np.linspace(-4, 4, 16)
         template = xr.DataArray(
             np.zeros((16, 16), dtype=np.float64),
             dims=['y', 'x'],
-            coords={'y': np.linspace(4, -4, 16), 'x': np.linspace(-4, 4, 16)},
+            coords={'y': ys, 'x': xs},
         )
         np_result = kde(x, y, bandwidth=1.0, kernel='quartic',
                         template=template)