Generalize local_subtensor_of_elemwise to Blockwise

Author: ricardoV94

pytensor/tensor/rewriting/subtensor_lift.py

@@ -20,6 +20,7 @@
     join,
     register_infer_shape,
 )
+from pytensor.tensor.blockwise import Blockwise
 from pytensor.tensor.elemwise import CAReduce, DimShuffle, Elemwise
 from pytensor.tensor.exceptions import NotScalarConstantError
 from pytensor.tensor.extra_ops import squeeze
@@ -169,16 +170,16 @@ def local_subtensor_of_dot(fgraph, node):
 @register_canonicalize("shape_unsafe")
 @register_specialize("shape_unsafe")
 @node_rewriter([Subtensor])
-def local_subtensor_of_elemwise(fgraph, node):
-    """Lift a Subtensor through an Elemwise and its implicit broadcasting behavior.
+def local_subtensor_of_batch_dims(fgraph, node):
+    """Lift a Subtensor through the batch dims of an (Elemwise or Blockwise) operation and its implicit broadcasting behavior.
 
     exp(x)[:, 0] -> exp(x[:, 0])
     add(x, y)[0] -> add(x[0], y[0])
     add(x[None], y)[2] -> add(x, y[2])
     """
     elem, *idx = node.inputs
 
-    if not (elem.owner and isinstance(elem.owner.op, Elemwise)):
+    if not (elem.owner and isinstance(elem.owner.op, Elemwise | Blockwise)):
         return None
 
     if len(fgraph.clients[elem]) > 1:
@@ -188,9 +189,34 @@ def local_subtensor_of_elemwise(fgraph, node):
 
     idx_tuple = indices_from_subtensor(idx, node.op.idx_list)
 
+    batch_ndim = (
+        elem.owner.op.batch_ndim(elem.owner)
+        if isinstance(elem.owner.op, Blockwise)
+        else elem.ndim
+    )
+
+    if len(idx_tuple) > batch_ndim:
+        # Indexing on core dimensions of Blockwise. We split the indices and lift the batch ones only
+        batch_indices, core_indices = idx_tuple[:batch_ndim], idx_tuple[batch_ndim:]
+        if all(is_full_slice(idx) for idx in batch_indices):
+            # No batch indices, nothing to do
+            return None
+        elem_with_batch_indices = elem[batch_indices]
+        [elem_with_batch_indices_lifted] = local_subtensor_of_batch_dims.transform(
+            fgraph, elem_with_batch_indices.owner
+        )
+        # Reapply the core_indices
+        core_ndim = elem.type.ndim - batch_ndim
+        # Number of batch dims may have changed with the lifting of indices, so we recompute
+        new_batch_ndim = elem_with_batch_indices_lifted.type.ndim - core_ndim
+        new_indices = (*(slice(None),) * new_batch_ndim, *core_indices)
+        new_elem = elem_with_batch_indices_lifted[new_indices]
+        copy_stack_trace(node.outputs[0], new_elem)
+        return [new_elem]
+
     elem_inputs = elem.owner.inputs
-    elem_bcast = elem.type.broadcastable
-    if all(inp.type.broadcastable == elem_bcast for inp in elem_inputs):
+    elem_bcast = elem.type.broadcastable[:batch_ndim]
+    if all(inp.type.broadcastable[:batch_ndim] == elem_bcast for inp in elem_inputs):
         # No need to worry about implicit broadcasting.
         indexed_inputs = [inp[idx_tuple] for inp in elem_inputs]
 
@@ -201,7 +227,7 @@ def local_subtensor_of_elemwise(fgraph, node):
             zip(
                 idx_tuple,
                 elem_bcast,
-                *(inp.type.broadcastable for inp in elem_inputs),
+                *(inp.type.broadcastable[:batch_ndim] for inp in elem_inputs),
                 # Indices can be shorter than input ndims
                 strict=False,
             )

tests/tensor/rewriting/test_subtensor_lift.py

@@ -37,14 +37,16 @@
     vector,
 )
 from pytensor.tensor.basic import MakeVector, concatenate, expand_dims, make_vector
+from pytensor.tensor.blockwise import Blockwise
 from pytensor.tensor.elemwise import DimShuffle, Elemwise
 from pytensor.tensor.math import sum as pt_sum
 from pytensor.tensor.rewriting.subtensor_lift import (
     local_subtensor_make_vector,
-    local_subtensor_of_elemwise,
+    local_subtensor_of_batch_dims,
     local_subtensor_shape_constant,
 )
 from pytensor.tensor.shape import SpecifyShape, _shape
+from pytensor.tensor.signal import convolve1d
 from pytensor.tensor.special import softmax
 from pytensor.tensor.subtensor import AdvancedSubtensor, Subtensor
 
@@ -58,7 +60,7 @@
 NO_OPTIMIZATION_MODE = Mode(linker="py", optimizer=None)
 
 
-class TestLocalSubtensorOfElemwise:
+class TestLocalSubtensorOfBatchDims:
     def test_unary_multiple_clients(self):
         # as test0, but we reuse the output of the elemwise
         # So we should not lift the subtensor
@@ -144,7 +146,7 @@ def test_multinary_multiple_clients(self):
             ),
         ],
     )
-    def test_local_subtensor_of_elemwise(self, original_fn, expected_fn):
+    def test_elemwise(self, original_fn, expected_fn):
         rng = np.random.default_rng(257)
         x = pt.matrix("x", shape=(5, 3))
         y = pt.matrix("y", shape=(5, 3))
@@ -163,19 +165,50 @@ def test_local_subtensor_of_elemwise(self, original_fn, expected_fn):
             out.eval({x: x_test, y: y_test}, **eval_kwargs),
         )
 
-    def test_local_subtensor_of_elemwise_multiple_clients(self):
+    def test_elemwise_multiple_clients(self):
         x = pt.matrix("x", shape=(5, 3))
         y = pt.matrix("y", shape=(5, 3))
         out1 = add(x, y)
         out2 = out1[0]
 
         # Rewrite should fail when another node uses out1 directly (in this case it's an extra output)
         fgraph = FunctionGraph([x, y], [out1, out2], clone=False)
-        assert local_subtensor_of_elemwise.transform(fgraph, out2.owner) is None
+        assert local_subtensor_of_batch_dims.transform(fgraph, out2.owner) is None
 
         # Otherwise it should work
         fgraph.remove_output(0)
-        assert local_subtensor_of_elemwise.transform(fgraph, out2.owner) is not None
+        assert local_subtensor_of_batch_dims.transform(fgraph, out2.owner) is not None
+
+    def test_blockwise(self):
+        x = tensor3("x", shape=(7, 5, 11))
+        y = tensor("y", shape=(7, 33))
+        out = convolve1d(x, y[:, None, :])
+        assert isinstance(out.owner.op, Blockwise)
+
+        out_sliced = out[2:][:, 3:]
+        rewritten_out_sliced = rewrite_graph(out_sliced)
+        assert equal_computations(
+            [rewritten_out_sliced], [convolve1d(x[2:, 3:], y[2:][:, None, :])]
+        )
+
+        rng = np.random.default_rng(191)
+        x_test = rng.normal(size=x.type.shape).astype(x.type.dtype)
+        y_test = rng.normal(size=y.type.shape).astype(y.type.dtype)
+        np.testing.assert_allclose(
+            rewritten_out_sliced.eval(
+                {x: x_test, y: y_test}, mode=NO_OPTIMIZATION_MODE
+            ),
+            out_sliced.eval({x: x_test, y: y_test}, mode=NO_OPTIMIZATION_MODE),
+        )
+
+        # Check slice on core dims
+        # Note: if we implement a rewrite on the core dims, this test should be changed for another Blockwise
+        # that has no such rewrite or one created just for testing purposes
+        out_sliced = out[2:][:, 0][:, 4:]
+        rewritten_out_sliced = rewrite_graph(out_sliced)
+        assert equal_computations(
+            [rewritten_out_sliced], [convolve1d(x[2:, 0], y[2:])[:, 4:]]
+        )
 
 
 @pytest.mark.parametrize(