Support more cases of multi-dimensional advanced indexing and updating in Numba

Extends pre-existing rewrite to ravel multiple integer indices, and to place them consecutively. The following cases should now be supported without object mode:
* Advanced integer indexing (not mixed with basic or boolean indexing) that do not require broadcasting of indices
* Consecutive advanced integer indexing updating (set/inc) (not mixed with basic or boolean indexing) that do not require broadcasting of indices or y.

Author: ricardoV94

pytensor/link/numba/dispatch/subtensor.py

@@ -150,7 +150,7 @@ def broadcasted_to(x_bcast: tuple[bool, ...], to_bcast: tuple[bool, ...]):
             for adv_idx in adv_idxs
         )
         # Must be consecutive
-        and not op.non_contiguous_adv_indexing(node)
+        and not op.non_consecutive_adv_indexing(node)
         # y in set/inc_subtensor cannot be broadcasted
         and (
             y is None

pytensor/tensor/rewriting/subtensor.py

@@ -2029,18 +2029,41 @@ def ravel_multidimensional_bool_idx(fgraph, node):
     return [copy_stack_trace(node.outputs[0], new_out)]
 
 
-@node_rewriter(tracks=[AdvancedSubtensor])
+@node_rewriter(tracks=[AdvancedSubtensor, AdvancedIncSubtensor])
 def ravel_multidimensional_int_idx(fgraph, node):
-    """Convert multidimensional integer indexing into equivalent vector integer index, supported by Numba
-
-    x[eye(3, dtype=int)] -> x[eye(3).ravel()].reshape((3, 3))
+    """Convert multidimensional integer indexing into equivalent consecutive vector integer index,
+    supported by Numba or by our specialized dispatchers
 
+        x[eye(3)] -> x[eye(3).ravel()].reshape((3, 3))
 
     NOTE: This is very similar to the rewrite `local_replace_AdvancedSubtensor` except it also handles non-full slices
 
-    x[eye(3, dtype=int), 2:] -> x[eye(3).ravel(), 2:].reshape((3, 3, ...)), where ... are the remaining output shapes
+        x[eye(3), 2:] -> x[eye(3).ravel(), 2:].reshape((3, 3, ...)), where ... are the remaining output shapes
+
+    It also handles multiple integer indices, but only if they don't broadcast
+
+        x[eye(3,), 2:, eye(3)] -> x[eye(3), eye(3), 2:].reshape((3, 3, ...)), where ... are the remaining output shapes
+
+    Also handles AdvancedIncSubtensor, but only if the advanced indices are consecutive and neither indices nor y broadcast
+
+        x[eye(3), 2:].set(y) -> x[eye(3).ravel(), 2:].set(y.reshape(-1, y.shape[1:]))
+
     """
-    x, *idxs = node.inputs
+    op = node.op
+    non_consecutive_adv_indexing = op.non_consecutive_adv_indexing(node)
+    is_inc_subtensor = isinstance(op, AdvancedIncSubtensor)
+
+    if is_inc_subtensor:
+        x, y, *idxs = node.inputs
+        # Inc/SetSubtensor is harder to reason about due to y
+        # We get out if it's broadcasting or if the advanced indices are non-consecutive
+        if non_consecutive_adv_indexing or (
+            y.type.broadcastable != x[tuple(idxs)].type.broadcastable
+        ):
+            return None
+
+    else:
+        x, *idxs = node.inputs
 
     if any(
         (
@@ -2049,50 +2072,104 @@ def ravel_multidimensional_int_idx(fgraph, node):
         )
         for idx in idxs
     ):
-        # Get out if there are any other advanced indexes or np.newaxis
+        # Get out if there are any other advanced indices or np.newaxis
         return None
 
-    int_idxs = [
+    int_idxs_and_pos = [
         (i, idx)
         for i, idx in enumerate(idxs)
         if (isinstance(idx.type, TensorType) and idx.dtype in integer_dtypes)
     ]
 
-    if len(int_idxs) != 1:
-        # Get out if there are no or multiple integer idxs
+    if not int_idxs_and_pos:
         return None
 
-    [(int_idx_pos, int_idx)] = int_idxs
-    if int_idx.type.ndim < 2:
-        # No need to do anything if it's a vector or scalar, as it's already supported by Numba
+    int_idxs_pos, int_idxs = zip(
+        *int_idxs_and_pos, strict=False
+    )  # strict=False because by definition it's true
+
+    first_int_idx_pos = int_idxs_pos[0]
+    first_int_idx = int_idxs[0]
+    first_int_idx_bcast = first_int_idx.type.broadcastable
+
+    if any(int_idx.type.broadcastable != first_int_idx_bcast for int_idx in int_idxs):
+        # We don't have a view-only broadcasting operation
+        # Explicitly broadcasting the indices can incur a memory / copy overhead
         return None
 
-    raveled_int_idx = int_idx.ravel()
-    new_idxs = list(idxs)
-    new_idxs[int_idx_pos] = raveled_int_idx
-    raveled_subtensor = x[tuple(new_idxs)]
-
-    # Reshape into correct shape
-    # Because we only allow one advanced indexing, the output dimension corresponding to the raveled integer indexing
-    # must match the input position. If there were multiple advanced indexes, this could have been forcefully moved to the front
-    raveled_shape = raveled_subtensor.shape
-    unraveled_shape = (
-        *raveled_shape[:int_idx_pos],
-        *int_idx.shape,
-        *raveled_shape[int_idx_pos + 1 :],
-    )
-    new_out = raveled_subtensor.reshape(unraveled_shape)
+    int_idxs_ndim = len(first_int_idx_bcast)
+    if (
+        int_idxs_ndim == 0
+    ):  # This should be a basic indexing operation, rewrite elsewhere
+        return None
+
+    int_idxs_need_raveling = int_idxs_ndim > 1
+    if not (int_idxs_need_raveling or non_consecutive_adv_indexing):
+        # Numba or our dispatch natively supports consecutive vector indices, nothing needs to be done
+        return None
+
+    # Reorder non-consecutive indices
+    if non_consecutive_adv_indexing:
+        assert not is_inc_subtensor  # Sanity check that we got out if this was the case
+        # This case works as if all the advanced indices were on the front
+        transposition = list(int_idxs_pos) + [
+            i for i in range(len(idxs)) if i not in int_idxs_pos
+        ]
+        idxs = tuple(idxs[a] for a in transposition)
+        x = x.transpose(transposition)
+        first_int_idx_pos = 0
+        del int_idxs_pos  # Make sure they are not wrongly used
+
+    # Ravel multidimensional indices
+    if int_idxs_need_raveling:
+        idxs = list(idxs)
+        for idx_pos, int_idx in enumerate(int_idxs, start=first_int_idx_pos):
+            idxs[idx_pos] = int_idx.ravel()
+
+    # Index with reordered and/or raveled indices
+    new_subtensor = x[tuple(idxs)]
+
+    if is_inc_subtensor:
+        int_idx_ndim = len(first_int_idx_bcast)
+        y_shape = tuple(y.shape)
+        y_raveled_shape = (
+            *y_shape[:first_int_idx_pos],
+            -1,
+            *y_shape[first_int_idx_pos + int_idx_ndim :],
+        )
+        y_raveled = y.reshape(y_raveled_shape)
+
+        new_out = inc_subtensor(
+            new_subtensor,
+            y_raveled,
+            set_instead_of_inc=op.set_instead_of_inc,
+            ignore_duplicates=op.ignore_duplicates,
+            inplace=op.inplace,
+        )
+
+    else:
+        # Unravel advanced indexing dimensions
+        raveled_shape = tuple(new_subtensor.shape)
+        unraveled_shape = (
+            *raveled_shape[:first_int_idx_pos],
+            *first_int_idx.shape,
+            *raveled_shape[first_int_idx_pos + 1 :],
+        )
+        new_out = new_subtensor.reshape(unraveled_shape)
+
     return [copy_stack_trace(node.outputs[0], new_out)]
 
 
 optdb["specialize"].register(
     ravel_multidimensional_bool_idx.__name__,
     ravel_multidimensional_bool_idx,
     "numba",
+    use_db_name_as_tag=False,  # Not included if only "specialize" is requested
 )
 
 optdb["specialize"].register(
     ravel_multidimensional_int_idx.__name__,
     ravel_multidimensional_int_idx,
     "numba",
+    use_db_name_as_tag=False,  # Not included if only "specialize" is requested
 )

pytensor/tensor/subtensor.py

@@ -1,5 +1,6 @@
 import logging
 import sys
+import warnings
 from collections.abc import Callable, Iterable
 from itertools import chain, groupby
 from textwrap import dedent
@@ -580,7 +581,7 @@ def group_indices(indices):
     return idx_groups
 
 
-def _non_contiguous_adv_indexing(indices) -> bool:
+def _non_consecutive_adv_indexing(indices) -> bool:
     """Check if the advanced indexing is non-contiguous (i.e., split by basic indexing)."""
     idx_groups = group_indices(indices)
     # This means that there are at least two groups of advanced indexing separated by basic indexing
@@ -611,7 +612,7 @@ def indexed_result_shape(array_shape, indices, indices_are_shapes=False):
     remaining_dims = range(pytensor.tensor.basic.get_vector_length(array_shape))
     idx_groups = group_indices(indices)
 
-    if _non_contiguous_adv_indexing(indices):
+    if _non_consecutive_adv_indexing(indices):
         # In this case NumPy places the advanced index groups in the front of the array
         # https://numpy.org/devdocs/user/basics.indexing.html#combining-advanced-and-basic-indexing
         idx_groups = sorted(idx_groups, key=lambda x: x[0])
@@ -2796,6 +2797,13 @@ def grad(self, inputs, grads):
 
     @staticmethod
     def non_contiguous_adv_indexing(node: Apply) -> bool:
+        warnings.warn(
+            "Method was renamed to `non_consecutive_adv_indexing`", FutureWarning
+        )
+        return AdvancedSubtensor.non_consecutive_adv_indexing(node)
+
+    @staticmethod
+    def non_consecutive_adv_indexing(node: Apply) -> bool:
         """
         Check if the advanced indexing is non-contiguous (i.e. interrupted by basic indexing).
 
@@ -2817,7 +2825,7 @@ def non_contiguous_adv_indexing(node: Apply) -> bool:
             True if the advanced indexing is non-contiguous, False otherwise.
         """
         _, *idxs = node.inputs
-        return _non_contiguous_adv_indexing(idxs)
+        return _non_consecutive_adv_indexing(idxs)
 
 
 advanced_subtensor = AdvancedSubtensor()
@@ -2835,7 +2843,7 @@ def vectorize_advanced_subtensor(op: AdvancedSubtensor, node, *batch_inputs):
         if isinstance(batch_idx, TensorVariable)
     )
 
-    if idxs_are_batched or (x_is_batched and op.non_contiguous_adv_indexing(node)):
+    if idxs_are_batched or (x_is_batched and op.non_consecutive_adv_indexing(node)):
         # Fallback to Blockwise if idxs are batched or if we have non contiguous advanced indexing
         # which would put the indexed results to the left of the batch dimensions!
         # TODO: Not all cases must be handled by Blockwise, but the logic is complex
@@ -2954,6 +2962,13 @@ def grad(self, inpt, output_gradients):
 
     @staticmethod
     def non_contiguous_adv_indexing(node: Apply) -> bool:
+        warnings.warn(
+            "Method was renamed to `non_consecutive_adv_indexing`", FutureWarning
+        )
+        return AdvancedIncSubtensor.non_consecutive_adv_indexing(node)
+
+    @staticmethod
+    def non_consecutive_adv_indexing(node: Apply) -> bool:
         """
         Check if the advanced indexing is non-contiguous (i.e. interrupted by basic indexing).
 
@@ -2975,7 +2990,7 @@ def non_contiguous_adv_indexing(node: Apply) -> bool:
             True if the advanced indexing is non-contiguous, False otherwise.
         """
         _, _, *idxs = node.inputs
-        return _non_contiguous_adv_indexing(idxs)
+        return _non_consecutive_adv_indexing(idxs)
 
 
 advanced_inc_subtensor = AdvancedIncSubtensor()

tests/link/numba/test_subtensor.py

@@ -81,11 +81,6 @@ def test_AdvancedSubtensor1_out_of_bounds():
             (np.array([True, False, False])),
             False,
         ),
-        (
-            pt.as_tensor(np.arange(3 * 4 * 5).reshape((3, 4, 5))),
-            ([1, 2], [2, 3]),
-            False,
-        ),
         # Single multidimensional indexing (supported after specialization rewrites)
         (
             as_tensor(np.arange(3 * 3).reshape((3, 3))),
@@ -117,6 +112,12 @@ def test_AdvancedSubtensor1_out_of_bounds():
             (slice(2, None), np.eye(3).astype(bool)),
             False,
         ),
+        # Multiple vector indexing (supported by our dispatcher)
+        (
+            pt.as_tensor(np.arange(3 * 4 * 5).reshape((3, 4, 5))),
+            ([1, 2], [2, 3]),
+            False,
+        ),
         (
             as_tensor(np.arange(3 * 4 * 5).reshape((3, 4, 5))),
             (slice(None), [1, 2], [3, 4]),
@@ -127,18 +128,35 @@ def test_AdvancedSubtensor1_out_of_bounds():
             ([1, 2], [3, 4], [5, 6]),
             False,
         ),
-        # Non-contiguous vector indexing, only supported in obj mode
+        # Non-consecutive vector indexing, supported by our dispatcher after rewriting
         (
             as_tensor(np.arange(3 * 4 * 5).reshape((3, 4, 5))),
             ([1, 2], slice(None), [3, 4]),
-            True,
+            False,
+        ),
+        # Multiple multidimensional integer indexing (supported by our dispatcher)
+        (
+            as_tensor(np.arange(3 * 4 * 5).reshape((3, 4, 5))),
+            ([[1, 2], [2, 1]], [[0, 0], [0, 0]]),
+            False,
+        ),
+        (
+            as_tensor(np.arange(2 * 3 * 4 * 5).reshape((2, 3, 4, 5))),
+            (slice(None), [[1, 2], [2, 1]], slice(None), [[0, 0], [0, 0]]),
+            False,
         ),
-        # >1d vector indexing, only supported in obj mode
+        # Multiple multidimensional indexing with broadcasting, only supported in obj mode
         (
             as_tensor(np.arange(3 * 4 * 5).reshape((3, 4, 5))),
             ([[1, 2], [2, 1]], [0, 0]),
             True,
         ),
+        # multiple multidimensional integer indexing mixed with basic indexing, only supported in obj mode
+        (
+            as_tensor(np.arange(3 * 4 * 5).reshape((3, 4, 5))),
+            ([[1, 2], [2, 1]], slice(1, None), [[0, 0], [0, 0]]),
+            True,
+        ),
     ],
 )
 @pytest.mark.filterwarnings("error")  # Raise if we did not expect objmode to be needed
@@ -297,15 +315,15 @@ def test_AdvancedIncSubtensor1(x, y, indices):
         (
             np.arange(3 * 4 * 5).reshape((3, 4, 5)),
             -np.arange(4 * 5).reshape(4, 5),
-            (0, [1, 2, 2, 3]),  # Broadcasted vector index
+            (0, [1, 2, 2, 3]),  # Broadcasted vector index with repeated values
             True,
             False,
             True,
         ),
         (
             np.arange(3 * 4 * 5).reshape((3, 4, 5)),
             np.array([-99]),  # Broadcasted value
-            (0, [1, 2, 2, 3]),  # Broadcasted vector index
+            (0, [1, 2, 2, 3]),  # Broadcasted vector index with repeated values
             True,
             False,
             True,
@@ -380,7 +398,7 @@ def test_AdvancedIncSubtensor1(x, y, indices):
         (
             np.arange(3 * 4 * 5).reshape((3, 4, 5)),
             rng.poisson(size=(2, 4)),
-            ([1, 2], slice(None), [3, 4]),  # Non-contiguous vector indices
+            ([1, 2], slice(None), [3, 4]),  # Non-consecutive vector indices
             False,
             True,
             True,
@@ -400,15 +418,23 @@ def test_AdvancedIncSubtensor1(x, y, indices):
         (
             np.arange(5),
             rng.poisson(size=(2, 2)),
-            ([[1, 2], [2, 3]]),  # matrix indices
+            ([[1, 2], [2, 3]]),  # matrix index
+            False,
+            False,
+            False,
+        ),
+        (
+            np.arange(3 * 5).reshape((3, 5)),
+            rng.poisson(size=(2, 2, 2)),
+            (slice(1, 3), [[1, 2], [2, 3]]),  # matrix index, mixed with basic index
+            False,
+            False,
             False,
-            False,  # Gets converted to AdvancedIncSubtensor1
-            True,  # This is actually supported with the default `ignore_duplicates=False`
         ),
         (
             np.arange(3 * 5).reshape((3, 5)),
-            rng.poisson(size=(1, 2, 2)),
-            (slice(1, 3), [[1, 2], [2, 3]]),  # matrix indices, mixed with basic index
+            rng.poisson(size=(1, 2, 2)),  # Same as before, but Y broadcasts
+            (slice(1, 3), [[1, 2], [2, 3]]),
             False,
             True,
             True,
@@ -421,6 +447,14 @@ def test_AdvancedIncSubtensor1(x, y, indices):
             False,
             False,
         ),
+        (
+            np.arange(3 * 4 * 5).reshape((3, 4, 5)),
+            rng.poisson(size=(3, 2, 2)),
+            (slice(None), [[1, 2], [2, 1]], [[2, 3], [0, 0]]),  # 2 matrix indices
+            False,
+            False,
+            False,
+        ),
     ],
 )
 @pytest.mark.parametrize("inplace", (False, True))