Forbid runtime broadcasting in Elemwise

Author: ricardoV94

pytensor/link/jax/dispatch/elemwise.py

@@ -7,9 +7,15 @@
 
 
 @jax_funcify.register(Elemwise)
-def jax_funcify_Elemwise(op, **kwargs):
+def jax_funcify_Elemwise(op, node, **kwargs):
     scalar_op = op.scalar_op
-    return jax_funcify(scalar_op, **kwargs)
+    base_fn = jax_funcify(scalar_op, node, **kwargs)
+
+    def elemwise_fn(*inputs):
+        Elemwise._check_runtime_broadcast(node, inputs)
+        return base_fn(*inputs)
+
+    return elemwise_fn
 
 
 @jax_funcify.register(CAReduce)

pytensor/tensor/__init__.py

@@ -132,6 +132,7 @@ def _get_vector_length_Constant(op: Union[Op, Variable], var: Constant) -> int:
     shape_padaxis,
     shape_padleft,
     shape_padright,
+    specify_broadcastable,
     specify_shape,
 )
 

pytensor/tensor/elemwise.py

@@ -6,7 +6,7 @@
 import pytensor.tensor.basic
 from pytensor.configdefaults import config
 from pytensor.gradient import DisconnectedType
-from pytensor.graph.basic import Apply
+from pytensor.graph.basic import Apply, Constant
 from pytensor.graph.null_type import NullType
 from pytensor.graph.utils import MethodNotDefined
 from pytensor.link.c.basic import failure_code
@@ -19,9 +19,9 @@
 from pytensor.scalar.basic import bool as scalar_bool
 from pytensor.scalar.basic import identity as scalar_identity
 from pytensor.scalar.basic import transfer_type, upcast
-from pytensor.tensor import _get_vector_length, as_tensor_variable
 from pytensor.tensor import elemwise_cgen as cgen
 from pytensor.tensor import get_vector_length
+from pytensor.tensor.basic import _get_vector_length, as_tensor_variable
 from pytensor.tensor.type import (
     TensorType,
     continuous_dtypes,
@@ -737,9 +737,7 @@ def perform(self, node, inputs, output_storage):
             # FIXME: This no longer calls the C implementation!
             super().perform(node, inputs, output_storage)
 
-        for d, dim_shapes in enumerate(zip(*(i.shape for i in inputs))):
-            if len(set(dim_shapes) - {1}) > 1:
-                raise ValueError(f"Shapes on dimension {d} do not match: {dim_shapes}")
+        self._check_runtime_broadcast(node, inputs)
 
         ufunc_args = inputs
         ufunc_kwargs = {}
@@ -815,18 +813,41 @@ def perform(self, node, inputs, output_storage):
             else:
                 storage[0] = variable
 
-    def infer_shape(self, fgraph, node, i_shapes) -> List[Tuple[TensorVariable, ...]]:
-        if len(node.outputs) > 1:
-            from pytensor.tensor.exceptions import ShapeError
-
-            raise ShapeError(
-                "Multiple outputs are not supported by the default `Elemwise.infer_shape`"
-            )
+    @staticmethod
+    def _check_runtime_broadcast(node, inputs):
+        for dims_and_bcast in zip(
+            *[
+                zip(input.shape, sinput.type.broadcastable)
+                for input, sinput in zip(inputs, node.inputs)
+            ]
+        ):
+            if any(d != 1 for d, _ in dims_and_bcast) and (1, False) in dims_and_bcast:
+                raise ValueError(
+                    "Runtime broadcasting not allowed. "
+                    "At least one input has a distinct dimension length of 1, but was not marked as broadcastable.\n"
+                    "If broadcasting was intended, use `specify_broadcastable` on the relevant input."
+                )
+                raise ValueError
 
-        out_shape = pytensor.tensor.broadcast_shape(*i_shapes, arrays_are_shapes=True)
+    def infer_shape(self, fgraph, node, i_shapes) -> List[Tuple[TensorVariable, ...]]:
+        one = pytensor.tensor.basic.constant(1, dtype="int64")
+        output_shape = []
+        for dim, broadcastable in enumerate(node.outputs[0].type.broadcastable):
+            out_dim_length = one
+            if not broadcastable:
+                # There must be some input that is not broadcastable in this dim
+                for inp_shape, inp_var in zip(i_shapes, node.inputs):
+                    if not inp_var.type.broadcastable[dim]:
+                        # Give preference to constant dims
+                        if isinstance(inp_shape[dim], Constant):
+                            out_dim_length = inp_shape[dim]
+                            break
+                        # If we haven't yet seen a non-broadcastable dim, use this one
+                        if out_dim_length is one:
+                            out_dim_length = inp_shape[dim]
+            output_shape.append(as_tensor_variable(out_dim_length, dtype="int64"))
 
-        # The `as_tensor_variable` should convert `ScalarType`s to `TensorType`s
-        return [tuple(as_tensor_variable(s) for s in out_shape)]
+        return [tuple(output_shape)] * len(node.outputs)
 
     def _c_all(self, node, nodename, inames, onames, sub):
         # Some `Op`s directly call `Elemwise._c_all` or `Elemwise.c_code`
@@ -1190,7 +1211,7 @@ def c_support_code_apply(self, node, nodename):
         return support_code
 
     def c_code_cache_version_apply(self, node):
-        version = [14]  # the version corresponding to the c code in this Op
+        version = [15]  # the version corresponding to the c code in this Op
 
         # now we insert versions for the ops on which we depend...
         scalar_node = Apply(

pytensor/tensor/elemwise_cgen.py

@@ -66,12 +66,10 @@ def make_checks(loop_orders, dtypes, sub):
             if index != "x":
                 # Initialize the variables associated to the jth loop
                 # jump = stride - adjust
-                # If the variable has size 1 in that dim, we set the stride to zero to
-                # emulate broadcasting
                 jump = f"({var}_stride{index}) - ({adjust})"
                 init += f"""
                 {var}_n{index} = PyArray_DIMS({var})[{index}];
-                {var}_stride{index} = ({var}_n{index} == 1)? 0 : PyArray_STRIDES({var})[{index}] / sizeof({dtype});
+                {var}_stride{index} = PyArray_STRIDES({var})[{index}] / sizeof({dtype});
                 {var}_jump{index}_{j} = {jump};
                 """
                 adjust = f"{var}_n{index}*{var}_stride{index}"
@@ -86,88 +84,73 @@ def make_checks(loop_orders, dtypes, sub):
     # This loop builds multiple if conditions to verify that the
     # dimensions of the inputs match, and the first one that is true
     # raises an informative error message
+
+    runtime_broadcast_error_msg = (
+        "Runtime broadcasting not allowed. "
+        "One input had a distinct dimension length of 1, but was not marked as broadcastable: "
+        "(input[%%i].shape[%%i] = %%lld, input[%%i].shape[%%i] = %%lld). "
+        "If broadcasting was intended, use `specify_broadcastable` on the relevant input."
+    )
+
     for matches in zip(*loop_orders):
         to_compare = [(j, x) for j, x in enumerate(matches) if x != "x"]
 
         # elements of to_compare are pairs ( input_variable_idx, input_variable_dim_idx )
         if len(to_compare) < 2:
             continue
 
-        # Find first dimension size that is != 1
-        jl, xl = to_compare[-1]
-        non1size_dim_check = f"""
-            npy_intp non1size_dim{xl};
-            non1size_dim{xl} = """
-        for j, x in to_compare[:-1]:
-            non1size_dim_check += f"(%(lv{j})s_n{x} != 1) ? %(lv{j})s_n{x} : "
-        non1size_dim_check += f"%(lv{jl})s_n{xl};"
-        check += non1size_dim_check
-
-        # Check the nonsize1 dims match
-        # TODO: This is a bit inefficient because we are comparing one dimension against itself
-        check += f"""
-            if (non1size_dim{xl} != 1)
-            {{
-        """
-        for j, x in to_compare:
+        j0, x0 = to_compare[0]
+        for j, x in to_compare[1:]:
             check += f"""
-                if ((%(lv{j})s_n{x} != non1size_dim{x}) && (%(lv{j})s_n{x} != 1))
+            if (%(lv{j0})s_n{x0} != %(lv{j})s_n{x})
+            {{
+                if (%(lv{j0})s_n{x0} == 1 || %(lv{j})s_n{x} == 1)
                 {{
-                    PyErr_Format(PyExc_ValueError, "Input dimension mismatch. One other input has shape[%%i] = %%lld, but input[%%i].shape[%%i] = %%lld.",
-                       {x},
-                       (long long int) non1size_dim{x},
+                    PyErr_Format(PyExc_ValueError, "{runtime_broadcast_error_msg}",
+                   {j0},
+                   {x0},
+                   (long long int) %(lv{j0})s_n{x0},
+                   {j},
+                   {x},
+                   (long long int) %(lv{j})s_n{x}
+                    );
+                }} else {{
+                    PyErr_Format(PyExc_ValueError, "Input dimension mismatch: (input[%%i].shape[%%i] = %%lld, input[%%i].shape[%%i] = %%lld)",
+                       {j0},
+                       {x0},
+                       (long long int) %(lv{j0})s_n{x0},
                        {j},
                        {x},
                        (long long int) %(lv{j})s_n{x}
                     );
-                    %(fail)s
                 }}
-            """
-        check += """
-            }
+                %(fail)s
+            }}
         """
 
     return init % sub + check % sub
 
 
-def compute_broadcast_dimensions(array_name: str, loop_orders, sub) -> str:
-    """Create c_code to compute broadcasted dimensions of multiple arrays, arising from
-    Elemwise operations.
+def compute_outputs_dims(array_name: str, loop_orders, sub) -> str:
+    """Create c_code to compute the output dimensions of an Elemwise operation.
 
     The code returned by this function populates the array `array_name`, but does not
     initialize it.
 
-    TODO: We can decide to either specialize C code even further given the input types
-    or make it general, regardless of whether static broadcastable information is given
+    Note: We could specialize C code even further with the known static output shapes
     """
     dims_c_code = ""
     for i, candidates in enumerate(zip(*loop_orders)):
-        # TODO: Are candidates always either "x" or "i"? If that's the case we can
-        # simplify some logic here (e.g., we don't need to track the `idx`).
-        nonx_candidates = tuple(
-            (idx, c) for idx, c in enumerate(candidates) if c != "x"
-        )
-
-        # All inputs are known to be broadcastable
-        if not nonx_candidates:
+        # Borrow the length of the first non-broadcastable input dimension
+        for j, candidate in enumerate(candidates):
+            if candidate != "x":
+                var = sub[f"lv{int(j)}"]
+                dims_c_code += f"{array_name}[{i}] = {var}_n{candidate};\n"
+                break
+        # If none is non-broadcastable, the output dimension has a length of 1
+        else:  # no-break
             dims_c_code += f"{array_name}[{i}] = 1;\n"
-            continue
-
-        # There is only one informative source of size
-        if len(nonx_candidates) == 1:
-            idx, candidate = nonx_candidates[0]
-            var = sub[f"lv{int(idx)}"]
-            dims_c_code += f"{array_name}[{i}] = {var}_n{candidate};\n"
-            continue
 
-        # In this case any non-size 1 variable will define the right size
-        dims_c_code += f"{array_name}[{i}] = "
-        for idx, candidate in nonx_candidates[:-1]:
-            var = sub[f"lv{int(idx)}"]
-            dims_c_code += f"({var}_n{candidate} != 1)? {var}_n{candidate}: "
-        idx, candidate = nonx_candidates[-1]
-        var = sub[f"lv{idx}"]
-        dims_c_code += f"{var}_n{candidate};\n"
     return dims_c_code
 
 
@@ -186,7 +169,7 @@ def make_alloc(loop_orders, dtype, sub, fortran="0"):
     if type.startswith("PYTENSOR_COMPLEX"):
         type = type.replace("PYTENSOR_COMPLEX", "NPY_COMPLEX")
     nd = len(loop_orders[0])
-    init_dims = compute_broadcast_dimensions("dims", loop_orders, sub)
+    init_dims = compute_outputs_dims("dims", loop_orders, sub)
 
     # TODO: it would be interesting to allocate the output in such a
     # way that its contiguous dimensions match one of the input's
@@ -359,7 +342,7 @@ def make_reordered_loop(
 
     # Get the (sorted) total number of iterations of each loop
     declare_totals = f"int init_totals[{nnested}];\n"
-    declare_totals += compute_broadcast_dimensions("init_totals", init_loop_orders, sub)
+    declare_totals += compute_outputs_dims("init_totals", init_loop_orders, sub)
 
     # Sort totals to match the new order that was computed by sorting
     # the loop vector. One integer variable per loop is declared.

tests/link/jax/test_elemwise.py

@@ -4,6 +4,7 @@
 
 import pytensor
 import pytensor.tensor as at
+from pytensor.compile import get_mode
 from pytensor.configdefaults import config
 from pytensor.graph.fg import FunctionGraph
 from pytensor.graph.op import get_test_value
@@ -14,6 +15,11 @@
 from pytensor.tensor.special import SoftmaxGrad, log_softmax, softmax
 from pytensor.tensor.type import matrix, tensor, vector
 from tests.link.jax.test_basic import compare_jax_and_py
+from tests.tensor.test_elemwise import TestElemwise
+
+
+def test_elemwise_runtime_broadcast_error():
+    TestElemwise.check_runtime_broadcast_error(get_mode("JAX"))
 
 
 def test_jax_Dimshuffle():

tests/tensor/test_elemwise.py

@@ -18,7 +18,6 @@
 from pytensor.tensor import as_tensor_variable
 from pytensor.tensor.basic import second
 from pytensor.tensor.elemwise import CAReduce, DimShuffle, Elemwise
-from pytensor.tensor.exceptions import ShapeError
 from pytensor.tensor.math import all as at_all
 from pytensor.tensor.math import any as at_any
 from pytensor.tensor.math import exp
@@ -769,10 +768,9 @@ def test_input_dimensions_overflow(self):
         g = pytensor.function([a, b, c, d, e, f], s, mode=Mode(linker="py"))
         g(*[np.zeros(2**11, config.floatX) for i in range(6)])
 
-    def check_input_dimensions_match(self, mode):
-        """Make sure that our input validation works correctly and doesn't
-        throw erroneous broadcast-based errors.
-        """
+    @staticmethod
+    def check_runtime_broadcast_error(mode):
+        """Check we emmit a clear error when runtime broadcasting would occur according to Numpy rules."""
         x_v = matrix("x")
         m_v = vector("m")
 
@@ -782,19 +780,18 @@ def check_input_dimensions_match(self, mode):
         z_v = x_v - m_v
         f = pytensor.function([x_v, m_v], z_v, mode=mode)
 
-        res = f(x, m)
+        with pytest.raises(ValueError, match="Runtime broadcasting not allowe"):
+            f(x, m)
 
-        assert np.array_equal(res, x - m)
-
-    def test_input_dimensions_match_python(self):
-        self.check_input_dimensions_match(Mode(linker="py"))
+    def test_runtime_broadcast_error_python(self):
+        self.check_runtime_broadcast_error(Mode(linker="py"))
 
     @pytest.mark.skipif(
         not pytensor.config.cxx,
         reason="G++ not available, so we need to skip this test.",
     )
-    def test_input_dimensions_match_c(self):
-        self.check_input_dimensions_match(Mode(linker="c"))
+    def test_runtime_broadcast_error_c(self):
+        self.check_runtime_broadcast_error(Mode(linker="c"))
 
     def test_str(self):
         op = Elemwise(aes.add, inplace_pattern={0: 0}, name=None)
@@ -819,7 +816,7 @@ def test_partial_static_shape_info(self):
         assert pytensor.get_underlying_scalar_constant(res_shape[0][0]) == 1
         assert pytensor.get_underlying_scalar_constant(res_shape[0][1]) == 1
 
-    def test_multi_output(self):
+    def test_infer_shape_multi_output(self):
         class CustomElemwise(Elemwise):
             def make_node(self, *args):
                 res = super().make_node(*args)
@@ -833,14 +830,26 @@ def make_node(self, *args):
                     ],
                 )
 
-        z_1, z_2 = CustomElemwise(aes.add)(
-            as_tensor_variable(np.eye(1)), as_tensor_variable(np.eye(1))
-        )
+        custom_elemwise = CustomElemwise(aes.add)
 
+        z_1, z_2 = custom_elemwise(
+            as_tensor_variable(np.eye(1)),
+            as_tensor_variable(np.eye(1)),
+        )
         in_1_shape = (aes.constant(1), aes.constant(1))
+        outs = z_1.owner.op.infer_shape(None, z_1.owner, [in_1_shape, in_1_shape])
+        for out in outs:
+            assert out[0].eval() == 1
+            assert out[1].eval() == 1
 
-        with pytest.raises(ShapeError):
-            z_1.owner.op.infer_shape(None, z_1.owner, [in_1_shape, in_1_shape])
+        z_1, z_2 = custom_elemwise(
+            as_tensor_variable(np.eye(1)), as_tensor_variable(np.eye(3))
+        )
+        in_2_shape = (aes.constant(3), aes.constant(3))
+        outs = z_1.owner.op.infer_shape(None, z_1.owner, [in_1_shape, in_2_shape])
+        for out in outs:
+            assert out[0].eval() == 3
+            assert out[1].eval() == 3
 
     def test_shape_types(self):
         x = tensor(dtype=np.float64, shape=(None, 1))