Speedup for bounding_box_query

src/spatialdata/_core/query/spatial_query.py

@@ -1,6 +1,5 @@
 from __future__ import annotations
 
-import warnings
 from abc import abstractmethod
 from collections.abc import Callable, Mapping
 from dataclasses import dataclass
@@ -9,6 +8,7 @@
 
 import dask.dataframe as dd
 import numpy as np
+import pandas as pd
 from dask.dataframe import DataFrame as DaskDataFrame
 from geopandas import GeoDataFrame
 from shapely.geometry import MultiPolygon, Point, Polygon
@@ -78,7 +78,7 @@ def _get_bounding_box_corners_in_intrinsic_coordinates(
 
     # compute the output axes of the transformation, remove c from input and output axes, return the matrix without c
     # and then build an affine transformation from that
-    m_without_c, input_axes_without_c, output_axes_without_c = _get_axes_of_tranformation(
+    m_without_c, input_axes_without_c, output_axes_without_c = _get_axes_of_transformation(
         element, target_coordinate_system
     )
     spatial_transform = Affine(m_without_c, input_axes=input_axes_without_c, output_axes=output_axes_without_c)
@@ -142,7 +142,7 @@ def _get_polygon_in_intrinsic_coordinates(
 
     polygon_gdf = ShapesModel.parse(GeoDataFrame(geometry=[polygon]))
 
-    m_without_c, input_axes_without_c, output_axes_without_c = _get_axes_of_tranformation(
+    m_without_c, input_axes_without_c, output_axes_without_c = _get_axes_of_transformation(
         element, target_coordinate_system
     )
     spatial_transform = Affine(m_without_c, input_axes=input_axes_without_c, output_axes=output_axes_without_c)
@@ -186,7 +186,7 @@ def _get_polygon_in_intrinsic_coordinates(
     return transform(polygon_gdf, to_coordinate_system="inverse")
 
 
-def _get_axes_of_tranformation(
+def _get_axes_of_transformation(
     element: SpatialElement, target_coordinate_system: str
 ) -> tuple[ArrayLike, tuple[str, ...], tuple[str, ...]]:
     """
@@ -321,6 +321,11 @@ def _get_case_of_bounding_box_query(
     return case
 
 
+def _is_scaling_transform(m_linear: np.ndarray) -> bool:
+    """True when the linear part is a diagonal (pure scaling) matrix."""
+    return np.allclose(m_linear, np.diag(np.diagonal(m_linear)))
+
+
 @dataclass(frozen=True)
 class BaseSpatialRequest:
     """Base class for spatial queries."""
@@ -386,6 +391,7 @@ def _bounding_box_mask_points(
     axes: tuple[str, ...],
     min_coordinate: list[Number] | ArrayLike,
     max_coordinate: list[Number] | ArrayLike,
+    points_df: pd.DataFrame | None = None,
 ) -> list[ArrayLike]:
     """Compute a mask that is true for the points inside axis-aligned bounding boxes.
 
@@ -404,31 +410,35 @@ def _bounding_box_mask_points(
         The lower right hand corners of the bounding boxes (i.e., the maximum coordinates along all dimensions).
         Shape: (n_boxes, n_axes) or (n_axes,) for a single box.
     {max_coordinate_docs}
+    points_df
+        A pre-computed pandas dataframe. Useful if the points_df has already been materialized, otherwise the methods simply
+        calls .compute() on the dask data frame
 
     Returns
     -------
     The masks for the points inside the bounding boxes.
     """
     element_axes = get_axes_names(points)
-
     min_coordinate = _parse_list_into_array(min_coordinate)
     max_coordinate = _parse_list_into_array(max_coordinate)
-
-    # Ensure min_coordinate and max_coordinate are 2D arrays
     min_coordinate = min_coordinate[np.newaxis, :] if min_coordinate.ndim == 1 else min_coordinate
     max_coordinate = max_coordinate[np.newaxis, :] if max_coordinate.ndim == 1 else max_coordinate
 
+    # Compute once here only if the caller hasn't already done so
+    if points_df is None:
+        points_df = points.compute()
+
     n_boxes = min_coordinate.shape[0]
     in_bounding_box_masks = []
-
     for box in range(n_boxes):
         box_masks = []
         for axis_index, axis_name in enumerate(axes):
             if axis_name not in element_axes:
                 continue
             min_value = min_coordinate[box, axis_index]
             max_value = max_coordinate[box, axis_index]
-            box_masks.append(points[axis_name].gt(min_value).compute() & points[axis_name].lt(max_value).compute())
+            col = points_df[axis_name].values
+            box_masks.append((col > min_value) & (col < max_value))
         bounding_box_mask = np.stack(box_masks, axis=-1)
         in_bounding_box_masks.append(np.all(bounding_box_mask, axis=1))
     return in_bounding_box_masks
@@ -514,16 +524,6 @@ def _(
     min_coordinate = _parse_list_into_array(min_coordinate)
     max_coordinate = _parse_list_into_array(max_coordinate)
     new_elements = {}
-    if sdata.points:
-        warnings.warn(
-            (
-                "The object has `points` element. Depending on the number of points, querying MAY suffer from "
-                "performance issues. Please consider filtering the object before calling this function by calling the "
-                "`subset()` method of `SpatialData`."
-            ),
-            UserWarning,
-            stacklevel=2,
-        )
     for element_type in ["points", "images", "labels", "shapes"]:
         elements = getattr(sdata, element_type)
         queried_elements = _dict_query_dispatcher(
@@ -630,7 +630,6 @@ def _(
     max_coordinate: list[Number] | ArrayLike,
     target_coordinate_system: str,
 ) -> DaskDataFrame | list[DaskDataFrame] | None:
-    from spatialdata import transform
     from spatialdata.transformations import get_transformation
 
     min_coordinate = _parse_list_into_array(min_coordinate)
@@ -648,100 +647,103 @@ def _(
         max_coordinate=max_coordinate,
     )
 
-    # get the four corners of the bounding box (2D case), or the 8 corners of the "3D bounding box" (3D case)
-    (intrinsic_bounding_box_corners, intrinsic_axes) = _get_bounding_box_corners_in_intrinsic_coordinates(
-        element=points,
-        axes=axes,
-        min_coordinate=min_coordinate,
-        max_coordinate=max_coordinate,
-        target_coordinate_system=target_coordinate_system,
+    m_without_c, input_axes_without_c, output_axes_without_c = _get_axes_of_transformation(
+        points, target_coordinate_system
+    )
+    m_without_c_linear = m_without_c[:-1, :-1]
+    _ = _get_case_of_bounding_box_query(
+        m_without_c_linear,
+        input_axes_without_c,
+        output_axes_without_c,
     )
-    min_coordinate_intrinsic = intrinsic_bounding_box_corners.min(dim="corner")
-    max_coordinate_intrinsic = intrinsic_bounding_box_corners.max(dim="corner")
-
-    min_coordinate_intrinsic = min_coordinate_intrinsic.data
-    max_coordinate_intrinsic = max_coordinate_intrinsic.data
-
-    # get the points in the intrinsic coordinate bounding box
-    in_intrinsic_bounding_box = _bounding_box_mask_points(
-        points=points,
-        axes=intrinsic_axes,
-        min_coordinate=min_coordinate_intrinsic,
-        max_coordinate=max_coordinate_intrinsic,
+    axes_adjusted, min_coordinate_adjusted, max_coordinate_adjusted = _adjust_bounding_box_to_real_axes(
+        axes,
+        min_coordinate,
+        max_coordinate,
+        output_axes_without_c,
     )
+    if axes_adjusted != output_axes_without_c:
+        raise RuntimeError("This should not happen")
 
-    if not (len_df := len(in_intrinsic_bounding_box)) == (len_bb := len(min_coordinate)):
-        raise ValueError(
-            f"Length of list of dataframes `{len_df}` is not equal to the number of bounding boxes axes `{len_bb}`."
-        )
-    points_in_intrinsic_bounding_box: list[DaskDataFrame | None] = []
+    # materialize the points in the intrinsic coordinate system once
     points_pd = points.compute()
-    attrs = points.attrs.copy()
-    for mask_np in in_intrinsic_bounding_box:
-        if mask_np.sum() == 0:
-            points_in_intrinsic_bounding_box.append(None)
-        else:
-            # TODO there is a problem when mixing dask dataframe graph with dask array graph. Need to compute for now.
-            # we can't compute either mask or points as when we calculate either one of them
-            # test_query_points_multiple_partitions will fail as the mask will be used to index each partition.
-            # However, if we compute and then create the dask array again we get the mixed dask graph problem.
-            filtered_pd = points_pd[mask_np]
-            points_filtered = dd.from_pandas(filtered_pd, npartitions=points.npartitions)
-            points_filtered.attrs.update(attrs)
-            points_in_intrinsic_bounding_box.append(points_filtered)
-    if len(points_in_intrinsic_bounding_box) == 0:
-        return None
 
-    # assert that the number of queried points is correct
-    assert len(points_in_intrinsic_bounding_box) == len(min_coordinate)
-
-    # # we have to reset the index since we have subset
-    # # https://stackoverflow.com/questions/61395351/how-to-reset-index-on-concatenated-dataframe-in-dask
-    # points_in_intrinsic_bounding_box = points_in_intrinsic_bounding_box.assign(idx=1)
-    # points_in_intrinsic_bounding_box = points_in_intrinsic_bounding_box.set_index(
-    #     points_in_intrinsic_bounding_box.idx.cumsum() - 1
-    # )
-    # points_in_intrinsic_bounding_box = points_in_intrinsic_bounding_box.map_partitions(
-    #     lambda df: df.rename(index={"idx": None})
-    # )
-    # points_in_intrinsic_bounding_box = points_in_intrinsic_bounding_box.drop(columns=["idx"])
-
-    # transform the element to the query coordinate system
+    # checking the type of the transformation
+    # in the case of an identity or scaling transform, we can skip the whole
+    # projection into intrinsic space and reprojection into the global coordinate system
+    is_identity_transform = input_axes_without_c == output_axes_without_c and np.allclose(
+        m_without_c, np.eye(m_without_c.shape[0])
+    )
+    is_scaling_transform = input_axes_without_c == output_axes_without_c and _is_scaling_transform(m_without_c_linear)
+
+    # if the transform is identity, we can save extra for the affine transformation
+    if is_identity_transform:
+        bounding_box_masks = _bounding_box_mask_points(
+            points=points,
+            axes=axes,
+            min_coordinate=min_coordinate,
+            max_coordinate=max_coordinate,
+            points_df=points_pd,
+        )
+    elif is_scaling_transform:
+        # Pull scale factors from the diagonal and the translation from the last column
+        scales = np.diagonal(m_without_c_linear)  # shape: (n_axes,)
+        translation = m_without_c[:-1, -1]  # shape: (n_axes,)
+
+        # Invert the affine: x_intrinsic = (x_output - translation) / scale
+        min_intrinsic = (min_coordinate_adjusted - translation) / scales
+        max_intrinsic = (max_coordinate_adjusted - translation) / scales
+
+        # Ensure min < max after inversion (negative scale flips the interval)
+        min_intrinsic, max_intrinsic = (
+            np.minimum(min_intrinsic, max_intrinsic),
+            np.maximum(min_intrinsic, max_intrinsic),
+        )
+
+        bounding_box_masks = _bounding_box_mask_points(
+            points=points,
+            axes=tuple(input_axes_without_c),
+            min_coordinate=min_intrinsic,
+            max_coordinate=max_intrinsic,
+            points_df=points_pd,
+        )
+    else:
+        query_coordinates = points_pd.loc[:, list(input_axes_without_c)].to_numpy(copy=False)
+        query_coordinates = query_coordinates @ m_without_c[:-1, :-1].T + m_without_c[:-1, -1]
+
+        bounding_box_masks = []
+        for box_index in range(min_coordinate_adjusted.shape[0]):
+            bounding_box_mask = np.ones(len(points_pd), dtype=bool)
+            for axis_index in range(len(output_axes_without_c)):
+                min_value = min_coordinate_adjusted[box_index, axis_index]
+                max_value = max_coordinate_adjusted[box_index, axis_index]
+                column = query_coordinates[:, axis_index]
+                bounding_box_mask &= (column > min_value) & (column < max_value)
+            bounding_box_masks.append(bounding_box_mask)
+
+    if not (len_df := len(bounding_box_masks)) == (len_bb := len(min_coordinate)):
+        raise ValueError(f"Length of list of masks `{len_df}` is not equal to the number of bounding boxes `{len_bb}`.")
+
+    old_transformations = get_transformation(points, get_all=True)
+    assert isinstance(old_transformations, dict)
+    feature_key = points.attrs.get(ATTRS_KEY, {}).get(PointsModel.FEATURE_KEY)
+
     output: list[DaskDataFrame | None] = []
-    for p, min_c, max_c in zip(points_in_intrinsic_bounding_box, min_coordinate, max_coordinate, strict=True):
-        if p is None:
+    for mask_np in bounding_box_masks:
+        bounding_box_indices = np.flatnonzero(mask_np)
+        if len(bounding_box_indices) == 0:
             output.append(None)
-        else:
-            points_query_coordinate_system = transform(
-                p, to_coordinate_system=target_coordinate_system, maintain_positioning=False
+            continue
+
+        # The exact mask is computed in the query coordinate system, but the returned points must stay intrinsic.
+        queried_points = points_pd.iloc[bounding_box_indices]
+        output.append(
+            PointsModel.parse(
+                dd.from_pandas(queried_points, npartitions=1),
+                transformations=old_transformations.copy(),
+                feature_key=feature_key,
             )
-
-            # get a mask for the points in the bounding box
-            bounding_box_mask = _bounding_box_mask_points(
-                points=points_query_coordinate_system,
-                axes=axes,
-                min_coordinate=min_c,  # type: ignore[arg-type]
-                max_coordinate=max_c,  # type: ignore[arg-type]
-            )
-            if len(bounding_box_mask) != 1:
-                raise ValueError(f"Expected a single mask, got {len(bounding_box_mask)} masks. Please report this bug.")
-            bounding_box_indices = np.where(bounding_box_mask[0])[0]
-
-            if len(bounding_box_indices) == 0:
-                output.append(None)
-            else:
-                points_df = p.compute().iloc[bounding_box_indices]
-                old_transformations = get_transformation(p, get_all=True)
-                assert isinstance(old_transformations, dict)
-                feature_key = p.attrs.get(ATTRS_KEY, {}).get(PointsModel.FEATURE_KEY)
-
-                output.append(
-                    PointsModel.parse(
-                        dd.from_pandas(points_df, npartitions=1),
-                        transformations=old_transformations.copy(),
-                        feature_key=feature_key,
-                    )
-                )
+        )
     if len(output) == 0:
         return None
     if len(output) == 1:
@@ -791,8 +793,8 @@ def _(
     )
     for box_corners in intrinsic_bounding_box_corners:
         bounding_box_non_axes_aligned = Polygon(box_corners.data)
-        indices = polygons.geometry.intersects(bounding_box_non_axes_aligned)
-        queried = polygons[indices]
+        candidate_idx = polygons.sindex.query(bounding_box_non_axes_aligned, predicate="intersects")
+        queried = polygons.iloc[candidate_idx]
         if len(queried) == 0:
             queried_polygon = None
         else:
@@ -949,17 +951,22 @@ def _(
         assert np.all(element[OLD_INDEX] == buffered.index)
     else:
         buffered[OLD_INDEX] = buffered.index
-    indices = buffered.geometry.apply(lambda x: x.intersects(polygon))
-    if np.sum(indices) == 0:
+
+    # Use sindex for fast candidate pre-filtering, then exact intersection check
+    # only on the (typically small) candidate set — same pattern as bounding_box_query.
+    candidate_idx = buffered.sindex.query(polygon, predicate="intersects")
+    if len(candidate_idx) == 0:
+        del buffered[OLD_INDEX]
         return None
-    queried_shapes = element[indices]
-    queried_shapes.index = buffered[indices][OLD_INDEX]
+
+    queried_shapes = element.iloc[candidate_idx].copy()
+    queried_shapes.index = buffered.iloc[candidate_idx][OLD_INDEX]
     queried_shapes.index.name = None
 
     if clip:
         if isinstance(element.geometry.iloc[0], Point):
-            queried_shapes = buffered[indices]
-            queried_shapes.index = buffered[indices][OLD_INDEX]
+            queried_shapes = buffered.iloc[candidate_idx].copy()
+            queried_shapes.index = buffered.iloc[candidate_idx][OLD_INDEX]
             queried_shapes.index.name = None
         queried_shapes = queried_shapes.clip(polygon_gdf, keep_geom_type=True)
 

tests/core/query/test_spatial_query.py

@@ -676,6 +676,35 @@ def _query(p: DaskDataFrame) -> DaskDataFrame:
     assert np.array_equal(q0.index.compute(), q1.index.compute())
 
 
+def test_query_points_multiple_boxes_in_transformed_coordinate_system():
+    from spatialdata.transformations import Affine
+
+    points_element = _make_points(np.array([[10, 10], [20, 30], [20, 30], [40, 50]]))
+    set_transformation(
+        points_element,
+        transformation=Affine(
+            np.array([[1, 0, 100], [0, 1, -50], [0, 0, 1]]),
+            input_axes=("x", "y"),
+            output_axes=("x", "y"),
+        ),
+        to_coordinate_system="aligned",
+    )
+
+    points_result = bounding_box_query(
+        points_element,
+        axes=("x", "y"),
+        min_coordinate=np.array([[118, -22], [138, -2], [200, 200]]),
+        max_coordinate=np.array([[122, -18], [142, 2], [210, 210]]),
+        target_coordinate_system="aligned",
+    )
+
+    np.testing.assert_allclose(points_result[0]["x"].compute(), [20, 20])
+    np.testing.assert_allclose(points_result[0]["y"].compute(), [30, 30])
+    np.testing.assert_allclose(points_result[1]["x"].compute(), [40])
+    np.testing.assert_allclose(points_result[1]["y"].compute(), [50])
+    assert points_result[2] is None
+
+
 @pytest.mark.parametrize("with_polygon_query", [True, False])
 @pytest.mark.parametrize(
     "name",