block_diag dot rewrite

Author: jessegrabowski

pytensor/tensor/rewriting/math.py

@@ -29,9 +29,11 @@
     constant,
     expand_dims,
     get_underlying_scalar_constant_value,
+    join,
     moveaxis,
     ones_like,
     register_infer_shape,
+    split,
     switch,
     zeros,
     zeros_like,
@@ -96,6 +98,7 @@
 from pytensor.tensor.rewriting.elemwise import apply_local_dimshuffle_lift
 from pytensor.tensor.rewriting.linalg import is_matrix_transpose
 from pytensor.tensor.shape import Shape, Shape_i
+from pytensor.tensor.slinalg import BlockDiagonal
 from pytensor.tensor.subtensor import Subtensor
 from pytensor.tensor.type import (
     complex_dtypes,
@@ -146,6 +149,72 @@ def local_0_dot_x(fgraph, node):
         return [zeros((x.shape[0], y.shape[1]), dtype=node.outputs[0].type.dtype)]
 
 
+@register_canonicalize
+@register_specialize
+@register_stabilize
+@node_rewriter([Dot])
+def local_block_diag_dot_to_dot_block_diag(fgraph, node):
+    r"""
+    Perform the rewrite ``dot(block_diag(A, B), C) -> block_diag(dot(A, C), dot(B, C))``
+
+    BlockDiag results in the creation of a matrix of shape ``(n1 * n2, m1 * m2)``. Because dot has complexity
+    of approximately O(n^3), it's always better to perform two dot products on the smaller matrices, rather than
+    a single dot on the larger matrix.
+    """
+    x, y = node.inputs
+    op = node.op
+
+    def check_for_block_diag(x):
+        return x.owner and (
+            isinstance(x.owner.op, BlockDiagonal)
+            or isinstance(x.owner.op, Blockwise)
+            and isinstance(x.owner.op.core_op, BlockDiagonal)
+        )
+
+    if not (check_for_block_diag(x) or check_for_block_diag(y)):
+        return None
+
+    # Case 1: Only one input is BlockDiagonal. In this case, multiply all components of the block-diagonal with the
+    # non-block diagonal, and return a new block diagonal
+    if check_for_block_diag(x) and not check_for_block_diag(y):
+        components = x.owner.inputs
+        y_splits = split(
+            y,
+            splits_size=[component.shape[-1] for component in components],
+            n_splits=len(components),
+        )
+        new_components = [
+            op(component, y_split) for component, y_split in zip(components, y_splits)
+        ]
+        new_output = join(0, *new_components)
+    elif not check_for_block_diag(x) and check_for_block_diag(y):
+        components = y.owner.inputs
+        new_components = [op(x, component) for component in components]
+        new_output = join(0, *new_components)
+
+    # Case 2: Both inputs are BlockDiagonal. Here we can proceed only if the static shapes are known and identical. In
+    # that case, blockdiag(a,b) @ blockdiag(c, d) = blockdiag(a @ c, b @ d), but this is not true in the general case
+    elif any(shape is None for shape in (*x.type.shape, *y.type.shape)):
+        return None
+    elif x.ndim == y.ndim and all(
+        x_shape == y_shape for x_shape, y_shape in zip(x.type.shape, y.type.shape)
+    ):
+        x_components = x.owner.inputs
+        y_components = y.owner.inputs
+
+        if len(x_components) != len(y_components):
+            return None
+
+        new_output = BlockDiagonal(len(x_components))(
+            *[op(x_comp, y_comp) for x_comp, y_comp in zip(x_components, y_components)]
+        )
+    else:
+        return None
+
+    copy_stack_trace(node.outputs[0], new_output)
+    return [new_output]
+
+
 @register_canonicalize
 @node_rewriter([Dot, _matmul])
 def local_lift_transpose_through_dot(fgraph, node):
@@ -2582,7 +2651,6 @@ def add_calculate(num, denum, aslist=False, out_type=None):
     name="add_canonizer_group",
 )
 
-
 register_canonicalize(local_add_canonizer, "shape_unsafe", name="local_add_canonizer")
 
 
@@ -3720,7 +3788,6 @@ def logmexpm1_to_log1mexp(fgraph, node):
 )
 register_stabilize(logdiffexp_to_log1mexpdiff, name="logdiffexp_to_log1mexpdiff")
 
-
 # log(sigmoid(x) / (1 - sigmoid(x))) -> x
 # i.e logit(sigmoid(x)) -> x
 local_logit_sigmoid = PatternNodeRewriter(
@@ -3734,7 +3801,6 @@ def logmexpm1_to_log1mexp(fgraph, node):
 register_canonicalize(local_logit_sigmoid)
 register_specialize(local_logit_sigmoid)
 
-
 # sigmoid(log(x / (1-x)) -> x
 # i.e., sigmoid(logit(x)) -> x
 local_sigmoid_logit = PatternNodeRewriter(
@@ -3775,7 +3841,6 @@ def local_useless_conj(fgraph, node):
 
 register_specialize(local_polygamma_to_tri_gamma)
 
-
 local_log_kv = PatternNodeRewriter(
     # Rewrite log(kv(v, x)) = log(kve(v, x) * exp(-x)) -> log(kve(v, x)) - x
     # During stabilize -x is converted to -1.0 * x

tests/tensor/rewriting/test_math.py

@@ -115,6 +115,7 @@
     simplify_mul,
 )
 from pytensor.tensor.shape import Reshape, Shape_i, SpecifyShape, specify_shape
+from pytensor.tensor.slinalg import BlockDiagonal
 from pytensor.tensor.type import (
     TensorType,
     cmatrix,
@@ -4745,3 +4746,75 @@ def test_local_dot_to_mul(batched, a_shape, b_shape):
         out.eval({a: a_test, b: b_test}, mode=test_mode),
         rewritten_out.eval({a: a_test, b: b_test}, mode=test_mode),
     )
+
+
+def test_local_block_diag_dot_to_dot_block_diag():
+    """
+    Test that dot(block_diag(x, y,), z) is rewritten to concat(dot(x, z[:n]), dot(y, z[n:]))
+    """
+    a = tensor("a", shape=(4, 2))
+    b = tensor("b", shape=(2, 4))
+    c = tensor("c", shape=(4, 4))
+    d = tensor("d", shape=(10,))
+
+    x = pt.linalg.block_diag(a, b, c)
+    out = x @ d
+
+    fn = pytensor.function([a, b, c, d], out)
+    assert not any(
+        isinstance(node, BlockDiagonal) for node in fn.maker.fgraph.toposort()
+    )
+
+    fn_expected = pytensor.function(
+        [a, b, c, d],
+        out,
+        mode=get_default_mode().excluding("local_block_diag_dot_to_dot_block_diag"),
+    )
+
+    rng = np.random.default_rng()
+    a_val = rng.normal(size=a.type.shape).astype(a.type.dtype)
+    b_val = rng.normal(size=b.type.shape).astype(b.type.dtype)
+    c_val = rng.normal(size=c.type.shape).astype(c.type.dtype)
+    d_val = rng.normal(size=d.type.shape).astype(d.type.dtype)
+
+    np.testing.assert_allclose(
+        fn(a_val, b_val, c_val, d_val),
+        fn_expected(a_val, b_val, c_val, d_val),
+        atol=1e-6 if config.floatX == "float32" else 1e-12,
+        rtol=1e-6 if config.floatX == "float32" else 1e-12,
+    )
+
+
+@pytest.mark.parametrize("rewrite", [True, False], ids=["rewrite", "no_rewrite"])
+@pytest.mark.parametrize("size", [10, 100, 1000], ids=["small", "medium", "large"])
+def test_block_diag_dot_to_dot_concat_benchmark(benchmark, size, rewrite):
+    rng = np.random.default_rng()
+    a_size = int(rng.uniform(0, size))
+    b_size = int(rng.uniform(0, size - a_size))
+    c_size = size - a_size - b_size
+
+    a = tensor("a", shape=(a_size, a_size))
+    b = tensor("b", shape=(b_size, b_size))
+    c = tensor("c", shape=(c_size, c_size))
+    d = tensor("d", shape=(size,))
+
+    x = pt.linalg.block_diag(a, b, c)
+    out = x @ d
+
+    mode = get_default_mode()
+    if not rewrite:
+        mode = mode.excluding("local_block_diag_dot_to_dot_block_diag")
+    fn = pytensor.function([a, b, c, d], out, mode=mode)
+
+    a_val = rng.normal(size=a.type.shape).astype(a.type.dtype)
+    b_val = rng.normal(size=b.type.shape).astype(b.type.dtype)
+    c_val = rng.normal(size=c.type.shape).astype(c.type.dtype)
+    d_val = rng.normal(size=d.type.shape).astype(d.type.dtype)
+
+    benchmark(
+        fn,
+        a_val,
+        b_val,
+        c_val,
+        d_val,
+    )