Support consecutive integer vector indexing in Numba backend

Author: ricardoV94

pytensor/link/numba/dispatch/subtensor.py

@@ -5,6 +5,7 @@
 from pytensor.link.numba.dispatch.basic import generate_fallback_impl, numba_njit
 from pytensor.link.utils import compile_function_src, unique_name_generator
 from pytensor.tensor import TensorType
+from pytensor.tensor.rewriting.subtensor import is_full_slice
 from pytensor.tensor.subtensor import (
     AdvancedIncSubtensor,
     AdvancedIncSubtensor1,
@@ -13,6 +14,7 @@
     IncSubtensor,
     Subtensor,
 )
+from pytensor.tensor.type_other import NoneTypeT, SliceType
 
 
 @numba_funcify.register(Subtensor)
@@ -104,18 +106,73 @@ def {function_name}({", ".join(input_names)}):
 @numba_funcify.register(AdvancedSubtensor)
 @numba_funcify.register(AdvancedIncSubtensor)
 def numba_funcify_AdvancedSubtensor(op, node, **kwargs):
-    idxs = node.inputs[1:] if isinstance(op, AdvancedSubtensor) else node.inputs[2:]
-    adv_idxs_dims = [
-        idx.type.ndim
+    if isinstance(op, AdvancedSubtensor):
+        x, y, idxs = node.inputs[0], None, node.inputs[1:]
+    else:
+        x, y, *idxs = node.inputs
+
+    basic_idxs = [
+        idx
         for idx in idxs
-        if (isinstance(idx.type, TensorType) and idx.type.ndim > 0)
+        if (
+            isinstance(idx.type, NoneTypeT)
+            or (isinstance(idx.type, SliceType) and not is_full_slice(idx))
+        )
+    ]
+    adv_idxs = [
+        {
+            "axis": i,
+            "dtype": idx.type.dtype,
+            "bcast": idx.type.broadcastable,
+            "ndim": idx.type.ndim,
+        }
+        for i, idx in enumerate(idxs)
+        if isinstance(idx.type, TensorType)
     ]
 
+    # Special case for consecutive consecutive vector indices
+    def broadcasted_to(x_bcast: tuple[bool, ...], to_bcast: tuple[bool, ...]):
+        # Check that x is not broadcasted to y based on broadcastable info
+        if len(x_bcast) < len(to_bcast):
+            return True
+        for x_bcast_dim, to_bcast_dim in zip(x_bcast, to_bcast, strict=True):
+            if x_bcast_dim and not to_bcast_dim:
+                return True
+        return False
+
+    # Special implementation for consecutive vector indices
     if (
+        not basic_idxs
+        and len(adv_idxs) >= 2
+        # Must be integer vectors
+        # Todo: we could allow shape=(1,) if this is the shape of x
+        and all(
+            (adv_idx["bcast"] == (False,) and adv_idx["dtype"] != "bool")
+            for adv_idx in adv_idxs
+        )
+        # Must be consecutive
+        and not op.non_contiguous_adv_indexing(node)
+        # y in set/inc_subtensor cannot be broadcasted
+        and (
+            y is None
+            or not broadcasted_to(
+                y.type.broadcastable,
+                (
+                    x.type.broadcastable[: adv_idxs[0]["axis"]]
+                    + x.type.broadcastable[adv_idxs[-1]["axis"] :]
+                ),
+            )
+        )
+    ):
+        return numba_funcify_multiple_vector_integer_indexing(op, node, **kwargs)
+
+    # Other cases not natively supported by Numba (fallback to obj-mode)
+    if not (
         # Numba does not support indexes with more than one dimension
+        any(idx["ndim"] > 1 for idx in adv_idxs)
         # Nor multiple vector indexes
-        (len(adv_idxs_dims) > 1 or adv_idxs_dims[0] > 1)
-        # The default index implementation does not handle duplicate indices correctly
+        or sum(idx["ndim"] > 0 for idx in adv_idxs) > 1
+        # The default PyTensor implementation does not handle duplicate indices correctly
         or (
             isinstance(op, AdvancedIncSubtensor)
             and not op.set_instead_of_inc
@@ -124,9 +181,91 @@ def numba_funcify_AdvancedSubtensor(op, node, **kwargs):
     ):
         return generate_fallback_impl(op, node, **kwargs)
 
+    # What's left should all be supported natively by numba
     return numba_funcify_default_subtensor(op, node, **kwargs)
 
 
+def numba_funcify_multiple_vector_integer_indexing(
+    op: AdvancedSubtensor | AdvancedIncSubtensor, node, **kwargs
+):
+    # Special-case implementation for multiple consecutive vector integer indices (and set/incsubtensor)
+    if isinstance(op, AdvancedSubtensor):
+        y, idxs = None, node.inputs[1:]
+    else:
+        y, *idxs = node.inputs[1:]
+
+    first_axis = next(
+        i for i, idx in enumerate(idxs) if isinstance(idx.type, TensorType)
+    )
+    try:
+        after_last_axis = next(
+            i
+            for i, idx in enumerate(idxs[first_axis:], start=first_axis)
+            if not isinstance(idx.type, TensorType)
+        )
+    except StopIteration:
+        after_last_axis = len(idxs)
+
+    if isinstance(op, AdvancedSubtensor):
+
+        @numba_njit
+        def advanced_subtensor_multiple_vector(x, *idxs):
+            none_slices = idxs[:first_axis]
+            vec_idxs = idxs[first_axis:after_last_axis]
+
+            x_shape = x.shape
+            idx_shape = vec_idxs[0].shape
+            shape_bef = x_shape[:first_axis]
+            shape_aft = x_shape[after_last_axis:]
+            out_shape = (*shape_bef, *idx_shape, *shape_aft)
+            out_buffer = np.empty(out_shape, dtype=x.dtype)
+            for i, scalar_idxs in enumerate(zip(*vec_idxs)):  # noqa: B905
+                out_buffer[(*none_slices, i)] = x[(*none_slices, *scalar_idxs)]
+            return out_buffer
+
+        return advanced_subtensor_multiple_vector
+
+    elif op.set_instead_of_inc:
+        inplace = op.inplace
+
+        @numba_njit
+        def advanced_set_subtensor_multiple_vector(x, y, *idxs):
+            vec_idxs = idxs[first_axis:after_last_axis]
+            x_shape = x.shape
+
+            if inplace:
+                out = x
+            else:
+                out = x.copy()
+
+            for outer in np.ndindex(x_shape[:first_axis]):
+                for i, scalar_idxs in enumerate(zip(*vec_idxs)):  # noqa: B905
+                    out[(*outer, *scalar_idxs)] = y[(*outer, i)]
+            return out
+
+        return advanced_set_subtensor_multiple_vector
+
+    else:
+        inplace = op.inplace
+
+        @numba_njit
+        def advanced_inc_subtensor_multiple_vector(x, y, *idxs):
+            vec_idxs = idxs[first_axis:after_last_axis]
+            x_shape = x.shape
+
+            if inplace:
+                out = x
+            else:
+                out = x.copy()
+
+            for outer in np.ndindex(x_shape[:first_axis]):
+                for i, scalar_idxs in enumerate(zip(*vec_idxs)):  # noqa: B905
+                    out[(*outer, *scalar_idxs)] += y[(*outer, i)]
+            return out
+
+        return advanced_inc_subtensor_multiple_vector
+
+
 @numba_funcify.register(AdvancedIncSubtensor1)
 def numba_funcify_AdvancedIncSubtensor1(op, node, **kwargs):
     inplace = op.inplace

pytensor/tensor/subtensor.py

@@ -2937,6 +2937,31 @@ def grad(self, inpt, output_gradients):
             gy = _sum_grad_over_bcasted_dims(y, gy)
         return [gx, gy] + [DisconnectedType()() for _ in idxs]
 
+    @staticmethod
+    def non_contiguous_adv_indexing(node: Apply) -> bool:
+        """
+        Check if the advanced indexing is non-contiguous (i.e. interrupted by basic indexing).
+
+        This function checks if the advanced indexing is non-contiguous,
+        in which case the advanced index dimensions are placed on the left of the
+        output array, regardless of their opriginal position.
+
+        See: https://numpy.org/doc/stable/user/basics.indexing.html#combining-advanced-and-basic-indexing
+
+
+        Parameters
+        ----------
+        node : Apply
+            The node of the AdvancedSubtensor operation.
+
+        Returns
+        -------
+        bool
+            True if the advanced indexing is non-contiguous, False otherwise.
+        """
+        _, _, *idxs = node.inputs
+        return _non_contiguous_adv_indexing(idxs)
+
 
 advanced_inc_subtensor = AdvancedIncSubtensor()
 advanced_set_subtensor = AdvancedIncSubtensor(set_instead_of_inc=True)

tests/link/numba/test_basic.py

@@ -228,9 +228,11 @@ def compare_numba_and_py(
     fgraph: FunctionGraph | tuple[Sequence["Variable"], Sequence["Variable"]],
     inputs: Sequence["TensorLike"],
     assert_fn: Callable | None = None,
+    *,
     numba_mode=numba_mode,
     py_mode=py_mode,
     updates=None,
+    inplace: bool = False,
     eval_obj_mode: bool = True,
 ) -> tuple[Callable, Any]:
     """Function to compare python graph output and Numba compiled output for testing equality
@@ -276,7 +278,14 @@ def assert_fn(x, y):
     pytensor_py_fn = function(
         fn_inputs, fn_outputs, mode=py_mode, accept_inplace=True, updates=updates
     )
-    py_res = pytensor_py_fn(*inputs)
+
+    test_inputs = (inp.copy() for inp in inputs) if inplace else inputs
+    py_res = pytensor_py_fn(*test_inputs)
+
+    # Get some coverage (and catch errors in python mode before unreadable numba ones)
+    if eval_obj_mode:
+        test_inputs = (inp.copy() for inp in inputs) if inplace else inputs
+        eval_python_only(fn_inputs, fn_outputs, test_inputs, mode=numba_mode)
 
     pytensor_numba_fn = function(
         fn_inputs,
@@ -285,11 +294,9 @@ def assert_fn(x, y):
         accept_inplace=True,
         updates=updates,
     )
-    numba_res = pytensor_numba_fn(*inputs)
 
-    # Get some coverage
-    if eval_obj_mode:
-        eval_python_only(fn_inputs, fn_outputs, inputs, mode=numba_mode)
+    test_inputs = (inp.copy() for inp in inputs) if inplace else inputs
+    numba_res = pytensor_numba_fn(*test_inputs)
 
     if len(fn_outputs) > 1:
         for j, p in zip(numba_res, py_res, strict=True):