Allow inplacing of SITSOT and last MITSOT in numba Scan, when they are discarded immediately

Author: ricardoV94

pytensor/link/numba/dispatch/scan.py

@@ -55,7 +55,7 @@ def range_arr(x):
 
 
 @numba_funcify.register(Scan)
-def numba_funcify_Scan(op, node, **kwargs):
+def numba_funcify_Scan(op: Scan, node, **kwargs):
     # Apply inner rewrites
     # TODO: Not sure this is the right place to do this, should we have a rewrite that
     #  explicitly triggers the optimization of the inner graphs of Scan?
@@ -67,9 +67,32 @@ def numba_funcify_Scan(op, node, **kwargs):
         .optimizer
     )
     fgraph = op.fgraph
+    # When the buffer can only hold one SITSOT or as as many MITSOT as there are taps,
+    # We must always discard the oldest tap, so it's safe to destroy it in the inner function.
+    # TODO: Allow inplace for MITMOT
+    destroyable_sitsot = [
+        inner_sitsot
+        for outer_sitsot, inner_sitsot in zip(
+            op.outer_sitsot(node.inputs), op.inner_sitsot(fgraph.inputs), strict=True
+        )
+        if outer_sitsot.type.shape[0] == 1
+    ]
+    destroyable_mitsot = [
+        oldest_inner_mitmot
+        for outer_mitsot, oldest_inner_mitmot, taps in zip(
+            op.outer_mitsot(node.inputs),
+            op.oldest_inner_mitsot(fgraph.inputs),
+            op.info.mit_sot_in_slices,
+            strict=True,
+        )
+        if outer_mitsot.type.shape[0] == abs(min(taps))
+    ]
+    destroyable = {*destroyable_sitsot, *destroyable_mitsot}
     add_supervisor_to_fgraph(
         fgraph=fgraph,
-        input_specs=[In(x, borrow=True, mutable=False) for x in fgraph.inputs],
+        input_specs=[
+            In(x, borrow=True, mutable=x in destroyable) for x in fgraph.inputs
+        ],
         accept_inplace=True,
     )
     rewriter(fgraph)

pytensor/scan/op.py

@@ -321,6 +321,16 @@ def inner_mitsot(self, list_inputs):
             self.info.n_seqs + n_mitmot_taps : self.info.n_seqs + ntaps_upto_sit_sot
         ]
 
+    def oldest_inner_mitsot(self, list_inputs):
+        inner_mitsot_inputs = self.inner_mitsot(list_inputs)
+        oldest_inner_mitsot_inputs = []
+        offset = 0
+        for taps in self.info.mit_sot_in_slices:
+            oldest_tap = np.argmin(taps)
+            oldest_inner_mitsot_inputs += [inner_mitsot_inputs[offset + oldest_tap]]
+            offset += len(taps)
+        return oldest_inner_mitsot_inputs
+
     def outer_mitsot(self, list_inputs):
         offset = 1 + self.info.n_seqs + self.info.n_mit_mot
         return list_inputs[offset : offset + self.info.n_mit_sot]

tests/link/numba/test_scan.py

@@ -451,6 +451,68 @@ def step(seq1, seq2, mitsot1, mitsot2, sitsot1):
     benchmark(numba_fn, *test.values())
 
 
+@pytest.mark.parametrize("n_steps_constant", (True, False))
+def test_inplace_taps(n_steps_constant):
+    """Test that numba will inplace in the inner_function of the oldest sit-sot, mit-sot taps."""
+    n_steps = 10 if n_steps_constant else scalar("n_steps", dtype=int)
+    a = scalar("a")
+    x0 = scalar("x0")
+    y0 = vector("y0", shape=(3,))
+
+    def step(xtm1, ytm1, ytm3, a):
+        x = xtm1 + 1
+        y = ytm1 + 1 + ytm3 + a
+        return x, x + y, y
+
+    [xs, zs, ys], _ = scan(
+        fn=step,
+        outputs_info=[
+            dict(initial=x0, taps=[-1]),
+            None,
+            dict(initial=y0, taps=[-1, -3]),
+        ],
+        non_sequences=[a],
+        n_steps=n_steps,
+    )
+    numba_fn, _ = compare_numba_and_py(
+        [n_steps] * (not n_steps_constant) + [a, x0, y0],
+        [xs[-1], zs[-1], ys[-1]],
+        [10] * (not n_steps_constant) + [np.pi, np.e, [0, np.euler_gamma, 1]],
+        numba_mode="NUMBA",
+        eval_obj_mode=False,
+    )
+    [scan_op] = [
+        node.op
+        for node in numba_fn.maker.fgraph.toposort()
+        if isinstance(node.op, Scan)
+    ]
+
+    # Scan reorders inputs internally, so we need to check its ordering
+    inner_inps = scan_op.fgraph.inputs
+    oldest_mit_sot_tap = scan_op.info.mit_sot_in_slices[0].index(-3)
+    oldest_mit_sot_inp = scan_op.inner_mitsot(inner_inps)[oldest_mit_sot_tap]
+    [sit_sot_inp] = scan_op.inner_sitsot(inner_inps)
+
+    inner_outs = scan_op.fgraph.outputs
+    [mit_sot_out] = scan_op.inner_mitsot_outs(inner_outs)
+    [sit_sot_out] = scan_op.inner_sitsot_outs(inner_outs)
+    [nit_sot_out] = scan_op.inner_nitsot_outs(inner_outs)
+
+    if n_steps_constant:
+        assert mit_sot_out.owner.op.destroy_map == {
+            0: [mit_sot_out.owner.inputs.index(oldest_mit_sot_inp)]
+        }
+        assert sit_sot_out.owner.op.destroy_map == {
+            0: [sit_sot_out.owner.inputs.index(sit_sot_inp)]
+        }
+    else:
+        # This is not a feature, but a current limitation
+        # https://github.com/pymc-devs/pytensor/issues/1283
+        assert mit_sot_out.owner.op.destroy_map == {}
+        assert sit_sot_out.owner.op.destroy_map == {}
+    assert nit_sot_out.owner.op.destroy_map == {}
+
+
 @pytest.mark.parametrize(
     "buffer_size", ("unit", "aligned", "misaligned", "whole", "whole+init")
 )