Numba UnravelIndex: Handle arbitrary indices ndim and F-order

Author: ricardoV94

pytensor/link/numba/dispatch/extra_ops.py

@@ -261,41 +261,60 @@ def unique(x):
 
 @register_funcify_and_cache_key(UnravelIndex)
 def numba_funcify_UnravelIndex(op, node, **kwargs):
-    order = op.order
-
-    if order != "C":
-        raise NotImplementedError(
-            "Numba does not support the `order` argument in `numpy.unravel_index`"
-        )
+    out_ndim = node.outputs[0].type.ndim
 
-    if len(node.outputs) == 1:
-
-        @numba_basic.numba_njit(inline="always")
-        def maybe_expand_dim(arr):
-            return arr
-
-    else:
+    if out_ndim == 0:
+        # Creating a tuple of 0d arrays in numba is basically impossible without codegen, so just go to obj_mode
+        return generate_fallback_impl(op, node=node), None
 
-        @numba_basic.numba_njit(inline="always")
-        def maybe_expand_dim(arr):
-            return np.expand_dims(arr, 1)
+    c_order = op.order == "C"
+    inp_ndim = node.inputs[0].type.ndim
+    transpose_axes = (inp_ndim, *range(inp_ndim))
 
     @numba_basic.numba_njit
-    def unravelindex(arr, shape):
+    def unravelindex(indices, shape):
         a = np.ones(len(shape), dtype=np.int64)
-        a[1:] = shape[:0:-1]
-        a = np.cumprod(a)[::-1]
+        if c_order:
+            # C-Order: Reverse shape (ignore dim0), cumulative product, then reverse back
+            # Strides: [dim1*dim2, dim2, 1]
+            a[1:] = shape[:0:-1]
+            a = np.cumprod(a)[::-1]
+        else:
+            # F-Order: Standard shape, cumulative product
+            # Strides: [1, dim0, dim0*dim1]
+            a[1:] = shape[:-1]
+            a = np.cumprod(a)
+
+        # Broadcast with a and shape on the last axis
+        unraveled_coords = (indices[..., None] // a) % shape
 
-        # PyTensor actually returns a `tuple` of these values, instead of an
-        # `ndarray`; however, this `ndarray` result should be able to be
-        # unpacked into a `tuple`, so this discrepancy shouldn't really matter
-        return ((maybe_expand_dim(arr) // a) % shape).T
+        # Then transpose it to the front
+        # Numba doesn't have moveaxis (why would it), so we use transpose
+        # res = np.moveaxis(res, -1, 0)
+        unraveled_coords = unraveled_coords.transpose(transpose_axes)
 
+        # This should be a tuple, but the array can be unpacked
+        # into multiple variables with the same effect by the outer function
+        # (special case for single entry is handled with an outer function below)
+        return unraveled_coords
+
+    cache_version = 1
     cache_key = sha256(
-        str((type(op), op.order, len(node.outputs))).encode()
+        str((type(op), op.order, len(node.outputs), cache_version)).encode()
     ).hexdigest()
 
-    return unravelindex, cache_key
+    if len(node.outputs) == 1:
+
+        @numba_basic.numba_njit
+        def unravel_index_single_item(arr, shape):
+            # Unpack single entry
+            (res,) = unravelindex(arr, shape)
+            return res
+
+        return unravel_index_single_item, cache_key
+
+    else:
+        return unravelindex, cache_key
 
 
 @register_funcify_default_op_cache_key(SearchsortedOp)

pytensor/tensor/extra_ops.py

@@ -1304,13 +1304,11 @@ def make_node(self, indices, dims):
         if dims.ndim != 1:
             raise TypeError("dims must be a 1D array")
 
+        out_type = indices.type.clone(dtype="int64")
         return Apply(
             self,
             [indices, dims],
-            [
-                TensorType(dtype="int64", shape=(None,) * indices.type.ndim)()
-                for i in range(ptb.get_vector_length(dims))
-            ],
+            [out_type() for _i in range(ptb.get_vector_length(dims))],
         )
 
     def infer_shape(self, fgraph, node, input_shapes):

tests/link/numba/test_extra_ops.py

@@ -1,4 +1,5 @@
 import contextlib
+from contextlib import nullcontext
 
 import numpy as np
 import pytest
@@ -295,37 +296,48 @@ def test_Unique(x, axis, return_index, return_inverse, return_counts, exc):
 
 
 @pytest.mark.parametrize(
-    "arr, shape, order, exc",
+    "arr, shape, requires_obj_mode",
     [
+        (
+            (pt.lscalar(), np.array(9, dtype="int64")),
+            pt.as_tensor([2, 3, 4]),
+            True,
+        ),
         (
             (pt.lvector(), np.array([9, 15, 1], dtype="int64")),
             pt.as_tensor([2, 3, 4]),
-            "C",
-            None,
+            False,
         ),
         (
             (pt.lvector(), np.array([1, 0], dtype="int64")),
             pt.as_tensor([2]),
-            "C",
-            None,
+            False,
         ),
         (
-            (pt.lvector(), np.array([9, 15, 1], dtype="int64")),
+            (pt.lmatrix(), np.array([[9, 15, 1], [1, 9, 15]], dtype="int64")),
             pt.as_tensor([2, 3, 4]),
-            "F",
-            NotImplementedError,
+            False,
         ),
     ],
 )
-def test_UnravelIndex(arr, shape, order, exc):
+def test_UnravelIndex(arr, shape, requires_obj_mode):
     arr, test_arr = arr
-    g = extra_ops.UnravelIndex(order)(arr, shape)
-
-    cm = contextlib.suppress() if exc is None else pytest.raises(exc)
+    g_c = extra_ops.UnravelIndex("C")(arr, shape)
+    g_f = extra_ops.UnravelIndex("F")(arr, shape)
+    if shape.type.shape == (1,):
+        outputs = [g_c, g_f]
+    else:
+        outputs = [*g_c, *g_f]
+
+    cm = (
+        pytest.warns(UserWarning, match="object mode")
+        if requires_obj_mode
+        else nullcontext()
+    )
     with cm:
         compare_numba_and_py(
             [arr],
-            g,
+            outputs,
             [test_arr],
         )