Benchmark Scan buffer optimization in Numba

Author: ricardoV94

tests/link/numba/test_scan.py

@@ -339,39 +339,6 @@ def power_step(prior_result, x):
     compare_numba_and_py([A], result, test_input_vals)
 
 
-@pytest.mark.parametrize("n_steps_val", [1, 5])
-def test_scan_save_mem_basic(n_steps_val):
-    """Make sure we can handle storage changes caused by the `scan_save_mem` rewrite."""
-
-    def f_pow2(x_tm2, x_tm1):
-        return 2 * x_tm1 + x_tm2
-
-    init_x = pt.dvector("init_x")
-    n_steps = pt.iscalar("n_steps")
-    output, _ = scan(
-        f_pow2,
-        sequences=[],
-        outputs_info=[{"initial": init_x, "taps": [-2, -1]}],
-        non_sequences=[],
-        n_steps=n_steps,
-    )
-
-    state_val = np.array([1.0, 2.0])
-
-    numba_mode = get_mode("NUMBA").including("scan_save_mem")
-    py_mode = Mode("py").including("scan_save_mem")
-
-    test_input_vals = (state_val, n_steps_val)
-
-    compare_numba_and_py(
-        [init_x, n_steps],
-        [output],
-        test_input_vals,
-        numba_mode=numba_mode,
-        py_mode=py_mode,
-    )
-
-
 def test_grad_sitsot():
     def get_sum_of_grad(inp):
         scan_outputs, updates = scan(
@@ -482,3 +449,120 @@ def step(seq1, seq2, mitsot1, mitsot2, sitsot1):
         np.testing.assert_array_almost_equal(numba_r, ref_r)
 
     benchmark(numba_fn, *test.values())
+
+
+@pytest.mark.parametrize(
+    "buffer_size", ("unit", "aligned", "misaligned", "whole", "whole+init")
+)
+@pytest.mark.parametrize("n_steps, op_size", [(10, 2), (512, 2), (512, 256)])
+class TestScanSITSOTBuffer:
+    def buffer_tester(self, n_steps, op_size, buffer_size, benchmark=None):
+        x0 = pt.vector(shape=(op_size,), dtype="float64")
+        xs, _ = pytensor.scan(
+            fn=lambda xtm1: (xtm1 + 1),
+            outputs_info=[x0],
+            n_steps=n_steps - 1,  # 1- makes it easier to align/misalign
+        )
+        if buffer_size == "unit":
+            xs_kept = xs[-1]  # Only last state is used
+            expected_buffer_size = 2
+        elif buffer_size == "aligned":
+            xs_kept = xs[-2:]  # The buffer will be aligned at the end of the 9 steps
+            expected_buffer_size = 2
+        elif buffer_size == "misaligned":
+            xs_kept = xs[-3:]  # The buffer will be misaligned at the end of the 9 steps
+            expected_buffer_size = 3
+        elif buffer_size == "whole":
+            xs_kept = xs  # What users think is the whole buffer
+            expected_buffer_size = n_steps - 1
+        elif buffer_size == "whole+init":
+            xs_kept = xs.owner.inputs[0]  # Whole buffer actually used by Scan
+            expected_buffer_size = n_steps
+
+        x_test = np.zeros(x0.type.shape)
+        numba_fn, _ = compare_numba_and_py(
+            [x0],
+            [xs_kept],
+            test_inputs=[x_test],
+            numba_mode="NUMBA",  # Default doesn't include optimizations
+            eval_obj_mode=False,
+        )
+        [scan_node] = [
+            node
+            for node in numba_fn.maker.fgraph.toposort()
+            if isinstance(node.op, Scan)
+        ]
+        buffer = scan_node.inputs[1]
+        assert buffer.type.shape[0] == expected_buffer_size
+
+        if benchmark is not None:
+            numba_fn.trust_input = True
+            benchmark(numba_fn, x_test)
+
+    def test_sit_sot_buffer(self, n_steps, op_size, buffer_size):
+        self.buffer_tester(n_steps, op_size, buffer_size, benchmark=None)
+
+    def test_sit_sot_buffer_benchmark(self, n_steps, op_size, buffer_size, benchmark):
+        self.buffer_tester(n_steps, op_size, buffer_size, benchmark=benchmark)
+
+
+@pytest.mark.parametrize("constant_n_steps", [False, True])
+@pytest.mark.parametrize("n_steps_val", [1, 1000])
+class TestScanMITSOTBuffer:
+    def buffer_tester(self, constant_n_steps, n_steps_val, benchmark=None):
+        """Make sure we can handle storage changes caused by the `scan_save_mem` rewrite."""
+
+        def f_pow2(x_tm2, x_tm1):
+            return 2 * x_tm1 + x_tm2
+
+        init_x = pt.vector("init_x", shape=(2,))
+        n_steps = pt.iscalar("n_steps")
+        output, _ = scan(
+            f_pow2,
+            sequences=[],
+            outputs_info=[{"initial": init_x, "taps": [-2, -1]}],
+            non_sequences=[],
+            n_steps=n_steps_val if constant_n_steps else n_steps,
+        )
+
+        init_x_val = np.array([1.0, 2.0], dtype=init_x.type.dtype)
+        test_vals = (
+            [init_x_val]
+            if constant_n_steps
+            else [init_x_val, np.asarray(n_steps_val, dtype=n_steps.type.dtype)]
+        )
+        numba_fn, _ = compare_numba_and_py(
+            [init_x] if constant_n_steps else [init_x, n_steps],
+            [output[-1]],
+            test_vals,
+            numba_mode="NUMBA",
+            eval_obj_mode=False,
+        )
+
+        if n_steps_val == 1 and constant_n_steps:
+            # There's no Scan in the graph when nsteps=constant(1)
+            return
+
+        # Check the buffer size as been optimized
+        [scan_node] = [
+            node
+            for node in numba_fn.maker.fgraph.toposort()
+            if isinstance(node.op, Scan)
+        ]
+        [mitsot_buffer] = scan_node.op.outer_mitsot(scan_node.inputs)
+        mitsot_buffer_shape = mitsot_buffer.shape.eval(
+            {init_x: init_x_val, n_steps: n_steps_val},
+            accept_inplace=True,
+            on_unused_input="ignore",
+        )
+        assert tuple(mitsot_buffer_shape) == (3,)
+
+        if benchmark is not None:
+            numba_fn.trust_input = True
+            benchmark(numba_fn, *test_vals)
+
+    def test_mit_sot_buffer(self, constant_n_steps, n_steps_val):
+        self.buffer_tester(constant_n_steps, n_steps_val, benchmark=None)
+
+    def test_mit_sot_buffer_benchmark(self, constant_n_steps, n_steps_val, benchmark):
+        self.buffer_tester(constant_n_steps, n_steps_val, benchmark=benchmark)