Rewrite solves involving kron to eliminate kron

Author: jessegrabowski

pytensor/tensor/rewriting/linalg.py

@@ -588,9 +588,11 @@ def svd_uv_merge(fgraph, node):
 @node_rewriter([Blockwise])
 def rewrite_inv_inv(fgraph, node):
     """
-    This rewrite takes advantage of the fact that if there are two consecutive inverse operations (inv(inv(input))), we get back our original input without having to compute inverse once.
+    This rewrite takes advantage of the fact that if there are two consecutive inverse operations (inv(inv(input))),
+    we get back our original input without having to compute inverse once.
 
-    Here, we check for direct inverse operations (inv/pinv)  and allows for any combination of these "inverse" nodes to be simply rewritten.
+    Here, we check for direct inverse operations (inv/pinv)  and allows for any combination of these "inverse" nodes to
+    be simply rewritten.
 
     Parameters
     ----------
@@ -855,6 +857,69 @@ def rewrite_det_kronecker(fgraph, node):
     return [det_final]
 
 
+@register_canonicalize
+@register_stabilize
+@node_rewriter([Blockwise])
+def rewrite_solve_kron_to_solve(fgraph, node):
+    """
+    Given a linear system of the form:
+
+    .. math::
+
+        (A \\otimes B) x = y
+
+    Define :math:`\text{vec}(x)` as a column-wise raveling operation (``x.reshape(-1, order='F')`` in code). Further,
+     define :math:`y = \text{vec}(Y)`. Then the above expression can be rewritten as:
+
+    .. math::
+
+        x = \text{vec}(B^{-1} Y A^{-T})
+
+    Eliminating the kronecker product from the expression.
+    """
+
+    if not isinstance(node.op.core_op, SolveBase):
+        return
+
+    solve_op = node.op
+    props_dict = solve_op.core_op._props_dict()
+    b_ndim = props_dict["b_ndim"]
+
+    A, b = node.inputs
+
+    if not A.owner or not (
+        isinstance(A.owner.op, KroneckerProduct)
+        or isinstance(A.owner.op, Blockwise)
+        and isinstance(A.owner.op.core_op, KroneckerProduct)
+    ):
+        return
+
+    x1, x2 = A.owner.inputs
+
+    m, n = x1.shape[-2], x2.shape[-2]
+    batch_shapes = x1.shape[:-2]
+
+    if b_ndim == 1:
+        # The rewritten expression will reshape B to be 2d. The easiest way to handle this is to just make a new
+        # solve node with n_ndim = 2
+        props_dict["b_ndim"] = 2
+        new_solve_op = Blockwise(type(solve_op.core_op)(**props_dict))
+        B = b.reshape((*batch_shapes, m, n))
+        res = new_solve_op(x1, new_solve_op(x2, B.mT).mT).reshape((*batch_shapes, -1))
+
+    else:
+        # If b_ndim is 2, we need to keep track of the original right-most dimension of b as an additional
+        # batch dimension
+        b_batch = b.shape[-1]
+        B = pt.moveaxis(b, -1, 0).reshape((b_batch, *batch_shapes, m, n))
+
+        res = pt.moveaxis(solve_op(x1, solve_op(x2, B.mT).mT), 0, -1).reshape(
+            (*batch_shapes, -1, b_batch)
+        )
+
+    return [res]
+
+
 @register_canonicalize
 @register_stabilize
 @node_rewriter([Blockwise])

tests/tensor/rewriting/test_linalg.py

@@ -828,6 +828,118 @@ def test_slogdet_kronecker_rewrite():
     )
 
 
+@pytest.mark.parametrize("add_batch", [True, False], ids=["batched", "not_batched"])
+@pytest.mark.parametrize("b_ndim", [1, 2], ids=["b_ndim_1", "b_ndim_2"])
+@pytest.mark.parametrize(
+    "solve_op, solve_kwargs",
+    [
+        (pt.linalg.solve, {"assume_a": "gen"}),
+        (pt.linalg.solve, {"assume_a": "pos"}),
+        (pt.linalg.solve, {"assume_a": "upper triangular"}),
+    ],
+    ids=["general", "positive definite", "triangular"],
+)
+def test_rewrite_solve_kron_to_solve(add_batch, b_ndim, solve_op, solve_kwargs):
+    a_shape = (5, 5) if not add_batch else (3, 5, 5)
+    b_shape = (5, 5) if not add_batch else (3, 5, 5)
+    A, B = pt.tensor("A", shape=a_shape), pt.tensor("B", shape=b_shape)
+
+    m, n = a_shape[-2], b_shape[-2]
+    y_shape = (m * n,)
+    if b_ndim == 2:
+        y_shape = (m * n, 5)
+    if add_batch:
+        y_shape = (3, *y_shape)
+
+    y = pt.tensor("y", shape=y_shape)
+    C = pt.vectorize(pt.linalg.kron, "(i,j),(k,l)->(m,n)")(A, B)
+
+    x = solve_op(C, y, **solve_kwargs, b_ndim=b_ndim)
+
+    def count_kron_ops(fn):
+        return sum(
+            [
+                isinstance(node.op, KroneckerProduct)
+                or (
+                    isinstance(node.op, Blockwise)
+                    and isinstance(node.op.core_op, KroneckerProduct)
+                )
+                for node in fn.maker.fgraph.apply_nodes
+            ]
+        )
+
+    fn_expected = pytensor.function(
+        [A, B, y], x, mode=get_default_mode().excluding("rewrite_solve_kron_to_solve")
+    )
+    assert count_kron_ops(fn_expected) == 1
+
+    fn = pytensor.function([A, B, y], x)
+    assert (
+        count_kron_ops(fn) == 0
+    ), "Rewrite did not apply, KroneckerProduct found in the graph"
+
+    rng = np.random.default_rng(sum(map(ord, "Go away Kron!")))
+    a_val = rng.normal(size=a_shape).astype(config.floatX)
+    b_val = rng.normal(size=b_shape).astype(config.floatX)
+    y_val = rng.normal(size=y_shape).astype(config.floatX)
+
+    if solve_kwargs["assume_a"] == "pos":
+        a_val = a_val @ a_val.mT
+        b_val = b_val @ b_val.mT
+    elif solve_kwargs["assume_a"] == "upper triangular":
+        a_idx = np.tril_indices(n=a_shape[-2], m=a_shape[-1], k=-1)
+        b_idx = np.tril_indices(n=b_shape[-2], m=b_shape[-1], k=-1)
+
+        if len(a_shape) > 2:
+            a_idx = (slice(None, None), *a_idx)
+        if len(b_shape) > 2:
+            b_idx = (slice(None, None), *b_idx)
+
+        a_val[a_idx] = 0
+        b_val[b_idx] = 0
+
+    expected = fn_expected(a_val, b_val, y_val)
+    result = fn(a_val, b_val, y_val)
+
+    np.testing.assert_allclose(
+        expected,
+        result,
+        atol=1e-8 if config.floatX == "float64" else 1e-5,
+        rtol=1e-8 if config.floatX == "float64" else 1e-5,
+    )
+
+
+@pytest.mark.parametrize(
+    "a_shape, b_shape",
+    [((5, 5), (5, 5)), ((50, 50), (50, 50)), ((100, 100), (100, 100))],
+    ids=["small", "medium", "large"],
+)
+@pytest.mark.parametrize("rewrite", [True, False], ids=["rewrite", "no_rewrite"])
+def test_rewrite_solve_kron_to_solve_benchmark(a_shape, b_shape, rewrite, benchmark):
+    A, B = pt.tensor("A", shape=a_shape), pt.tensor("B", shape=b_shape)
+    C = pt.linalg.kron(A, B)
+
+    m, n = a_shape[-2], b_shape[-2]
+    has_batch = len(a_shape) == 3
+    y_shape = (a_shape[0], m * n) if has_batch else (m * n,)
+    y = pt.tensor("y", shape=y_shape)
+    x = pt.linalg.solve(C, y, b_ndim=1)
+
+    rng = np.random.default_rng(sum(map(ord, "Go away Kron!")))
+    a_val = rng.normal(size=a_shape).astype(config.floatX)
+    b_val = rng.normal(size=b_shape).astype(config.floatX)
+    y_val = rng.normal(size=y_shape).astype(config.floatX)
+
+    mode = (
+        get_default_mode()
+        if rewrite
+        else get_default_mode().excluding("rewrite_solve_kron_to_solve")
+    )
+
+    fn = pytensor.function([A, B, y], x, mode=mode)
+    benchmark(fn, a_val, b_val, y_val)
+
+
 def test_cholesky_eye_rewrite():
     x = pt.eye(10)
     L = pt.linalg.cholesky(x)