improves dataloader performance

src/spatialdata/_core/query/_utils.py

@@ -2,13 +2,22 @@
 
 from typing import Any
 
+import numba as nb
+import numpy as np
 from anndata import AnnData
+from datatree import DataTree
 from xarray import DataArray
 
 from spatialdata._core._elements import Tables
 from spatialdata._core.spatialdata import SpatialData
 from spatialdata._types import ArrayLike
 from spatialdata._utils import Number, _parse_list_into_array
+from spatialdata.transformations._utils import compute_coordinates
+from spatialdata.transformations.transformations import (
+    BaseTransformation,
+    Sequence,
+    Translation,
+)
 
 
 def get_bounding_box_corners(
@@ -36,37 +45,146 @@ def get_bounding_box_corners(
     min_coordinate = _parse_list_into_array(min_coordinate)
     max_coordinate = _parse_list_into_array(max_coordinate)
 
-    if len(min_coordinate) not in (2, 3):
+    if min_coordinate.ndim == 1:
+        min_coordinate = min_coordinate[np.newaxis, :]
+        max_coordinate = max_coordinate[np.newaxis, :]
+
+    if min_coordinate.shape[1] not in (2, 3):
         raise ValueError("bounding box must be 2D or 3D")
 
-    if len(min_coordinate) == 2:
+    num_boxes = min_coordinate.shape[0]
+    num_dims = min_coordinate.shape[1]
+
+    if num_dims == 2:
         # 2D bounding box
         assert len(axes) == 2
-        return DataArray(
+        corners = np.array(
             [
-                [min_coordinate[0], min_coordinate[1]],
-                [min_coordinate[0], max_coordinate[1]],
-                [max_coordinate[0], max_coordinate[1]],
-                [max_coordinate[0], min_coordinate[1]],
-            ],
-            coords={"corner": range(4), "axis": list(axes)},
+                [min_coordinate[:, 0], min_coordinate[:, 1]],
+                [min_coordinate[:, 0], max_coordinate[:, 1]],
+                [max_coordinate[:, 0], max_coordinate[:, 1]],
+                [max_coordinate[:, 0], min_coordinate[:, 1]],
+            ]
         )
-
-    # 3D bounding cube
-    assert len(axes) == 3
-    return DataArray(
-        [
-            [min_coordinate[0], min_coordinate[1], min_coordinate[2]],
-            [min_coordinate[0], min_coordinate[1], max_coordinate[2]],
-            [min_coordinate[0], max_coordinate[1], max_coordinate[2]],
-            [min_coordinate[0], max_coordinate[1], min_coordinate[2]],
-            [max_coordinate[0], min_coordinate[1], min_coordinate[2]],
-            [max_coordinate[0], min_coordinate[1], max_coordinate[2]],
-            [max_coordinate[0], max_coordinate[1], max_coordinate[2]],
-            [max_coordinate[0], max_coordinate[1], min_coordinate[2]],
-        ],
-        coords={"corner": range(8), "axis": list(axes)},
+        corners = np.transpose(corners, (2, 0, 1))
+    else:
+        # 3D bounding cube
+        assert len(axes) == 3
+        corners = np.array(
+            [
+                [min_coordinate[:, 0], min_coordinate[:, 1], min_coordinate[:, 2]],
+                [min_coordinate[:, 0], min_coordinate[:, 1], max_coordinate[:, 2]],
+                [min_coordinate[:, 0], max_coordinate[:, 1], max_coordinate[:, 2]],
+                [min_coordinate[:, 0], max_coordinate[:, 1], min_coordinate[:, 2]],
+                [max_coordinate[:, 0], min_coordinate[:, 1], min_coordinate[:, 2]],
+                [max_coordinate[:, 0], min_coordinate[:, 1], max_coordinate[:, 2]],
+                [max_coordinate[:, 0], max_coordinate[:, 1], max_coordinate[:, 2]],
+                [max_coordinate[:, 0], max_coordinate[:, 1], min_coordinate[:, 2]],
+            ]
+        )
+        corners = np.transpose(corners, (2, 0, 1))
+    output = DataArray(
+        corners,
+        coords={
+            "box": range(num_boxes),
+            "corner": range(corners.shape[1]),
+            "axis": list(axes),
+        },
     )
+    if num_boxes > 1:
+        return output
+    return output.squeeze().drop_vars("box")
+
+
+@nb.njit(parallel=False, nopython=True)
+def _create_slices_and_translation(
+    min_values: nb.types.Array[nb.float64, nb.float64],
+    max_values: nb.types.Array[nb.float64, nb.float64],
+) -> tuple[nb.types.Array[nb.float64, nb.float64], nb.types.Array[nb.float64, nb.float64]]:
+    n_boxes, n_dims = min_values.shape
+    slices = np.empty((n_boxes, n_dims, 2), dtype=np.float64)  # (n_boxes, n_dims, [min, max])
+    translation_vectors = np.empty((n_boxes, n_dims), dtype=np.float64)  # (n_boxes, n_dims)
+
+    for i in range(n_boxes):
+        for j in range(n_dims):
+            slices[i, j, 0] = min_values[i, j]
+            slices[i, j, 1] = max_values[i, j]
+            translation_vectors[i, j] = np.ceil(max(min_values[i, j], 0))
+
+    return slices, translation_vectors
+
+
+def _process_data_tree_query_result(query_result: DataTree) -> DataTree | None:
+    d = {}
+    for k, data_tree in query_result.items():
+        v = data_tree.values()
+        assert len(v) == 1
+        xdata = v.__iter__().__next__()
+        if 0 in xdata.shape:
+            if k == "scale0":
+                return None
+        else:
+            d[k] = xdata
+
+    # Remove scales after finding a missing scale
+    scales_to_keep = []
+    for i, scale_name in enumerate(d.keys()):
+        if scale_name == f"scale{i}":
+            scales_to_keep.append(scale_name)
+        else:
+            break
+
+    # Case in which scale0 is not present but other scales are
+    if len(scales_to_keep) == 0:
+        return None
+
+    d = {k: d[k] for k in scales_to_keep}
+    result = DataTree.from_dict(d)
+
+    # Rechunk the data to avoid irregular chunks
+    for scale in result:
+        result[scale]["image"] = result[scale]["image"].chunk("auto")
+
+    return result
+
+
+def _process_query_result(
+    result: DataArray | DataTree, translation_vector: ArrayLike, axes: tuple[str, ...]
+) -> DataArray | DataTree | None:
+    from spatialdata.transformations import get_transformation, set_transformation
+
+    if isinstance(result, DataArray):
+        if 0 in result.shape:
+            return None
+        # rechunk the data to avoid irregular chunks
+        result = result.chunk("auto")
+    elif isinstance(result, DataTree):
+        result = _process_data_tree_query_result(result)
+        if result is None:
+            return None
+
+    result = compute_coordinates(result)
+
+    if not np.allclose(np.array(translation_vector), 0):
+        translation_transform = Translation(translation=translation_vector, axes=axes)
+
+        transformations = get_transformation(result, get_all=True)
+        assert isinstance(transformations, dict)
+
+        new_transformations = {}
+        for coordinate_system, initial_transform in transformations.items():
+            new_transformation: BaseTransformation = Sequence(
+                [translation_transform, initial_transform],
+            )
+            new_transformations[coordinate_system] = new_transformation
+        set_transformation(result, new_transformations, set_all=True)
+
+    # let's make a copy of the transformations so that we don't modify the original object
+    t = get_transformation(result, get_all=True)
+    assert isinstance(t, dict)
+    set_transformation(result, t.copy(), set_all=True)
+
+    return result
 
 
 def _get_filtered_or_unfiltered_tables(