Implement Blockwise Op to vectorize existing Ops

Inspired by: aesara-devs/aesara#1215

Co-authored-by: Brandon T. Willard <[email protected]>
Co-authored-by: Purna Chandra Mansingh <[email protected]>
Co-authored-by: Sayam Kumar <[email protected]>
Co-authored-by: Kaustubh <[email protected]>

Author: 5 people

pytensor/tensor/blockwise.py

@@ -22,13 +22,15 @@
 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.blockwise import _vectorize_node, vectorize_not_needed
 from pytensor.tensor.type import (
     TensorType,
     continuous_dtypes,
     discrete_dtypes,
     float_dtypes,
     lvector,
 )
+from pytensor.tensor.utils import broadcast_static_dim_lengths, import_func_from_string
 from pytensor.tensor.var import TensorVariable
 from pytensor.utils import uniq
 
@@ -232,7 +234,7 @@ def __str__(self):
             return f"Transpose{{axes={self.shuffle}}}"
         return f"DimShuffle{{order=[{','.join(map(str, self.new_order))}]}}"
 
-    def perform(self, node, inp, out, params):
+    def perform(self, node, inp, out, params=None):
         (res,) = inp
         (storage,) = out
 
@@ -429,28 +431,12 @@ def get_output_info(self, dim_shuffle, *inputs):
         # of all inputs in parallel... the all() gives us each output
         # broadcastable bit in turn.
 
-        def get_most_specialized_shape(shapes):
-            shapes = set(shapes)
-            # All shapes are the same
-            if len(shapes) == 1:
-                return tuple(shapes)[0]
-
-            # Only valid indeterminate case
-            if shapes == {None, 1}:
-                return None
-
-            shapes.discard(1)
-            shapes.discard(None)
-            if len(shapes) > 1:
-                raise ValueError
-            return tuple(shapes)[0]
-
         # it is multiplied by nout because Elemwise supports multiple outputs
         # (nout of them)
         try:
             out_shapes = [
                 [
-                    get_most_specialized_shape(shape)
+                    broadcast_static_dim_lengths(shape)
                     for shape in zip(*[inp.type.shape for inp in inputs])
                 ]
             ] * shadow.nout
@@ -665,22 +651,7 @@ def prepare_node(self, node, storage_map, compute_map, impl):
             impl = "c"
 
         if getattr(self, "nfunc_spec", None) and impl != "c":
-            self.nfunc = getattr(np, self.nfunc_spec[0], None)
-            if self.nfunc is None:
-                # Not inside NumPy. So probably another package like scipy.
-                symb = self.nfunc_spec[0].split(".")
-                for idx in range(1, len(self.nfunc_spec[0])):
-                    try:
-                        module = __import__(".".join(symb[:idx]))
-                    except ImportError:
-                        break
-                for sub in symb[1:]:
-                    try:
-                        module = getattr(module, sub)
-                    except AttributeError:
-                        module = None
-                        break
-                self.nfunc = module
+            self.nfunc = import_func_from_string(self.nfunc_spec[0])
 
         if (
             (len(node.inputs) + len(node.outputs)) <= 32
@@ -1768,3 +1739,37 @@ def _get_vector_length_Elemwise(op, var):
         return get_vector_length(var.owner.inputs[0])
 
     raise ValueError(f"Length of {var} cannot be determined")
+
+
+_vectorize_node.register(Elemwise, vectorize_not_needed)
+
+
+@_vectorize_node.register(DimShuffle)
+def vectorize_dimshuffle(op: DimShuffle, node: Apply, x: TensorVariable) -> Apply:
+    batched_ndims = x.type.ndim - node.inputs[0].type.ndim
+    if not batched_ndims:
+        return node.op.make_node(x)
+    input_broadcastable = x.type.broadcastable[:batched_ndims] + op.input_broadcastable
+    # e.g., ds(matrix, order=(1, "x", 0)) -> ds(tensor4, order=(0, 1, 3, "x", 2))
+    # e.g., ds(row, order=(1, "x")) -> ds(tensor4, order=(0, 1, 3, "x"))
+    new_order = list(range(batched_ndims)) + [
+        "x" if (o == "x") else (o + batched_ndims) for o in op.new_order
+    ]
+    return DimShuffle(input_broadcastable, new_order).make_node(x)
+
+
+@_vectorize_node.register(CAReduce)
+def vectorize_careduce(op: CAReduce, node: Apply, x: TensorVariable) -> Apply:
+    batched_ndims = x.type.ndim - node.inputs[0].type.ndim
+    if not batched_ndims:
+        return node.op.make_node(x)
+    axes = op.axis
+    # e.g., sum(matrix, axis=None) -> sum(tensor4, axis=(2, 3))
+    # e.g., sum(matrix, axis=0) -> sum(tensor4, axis=(2,))
+    if axes is None:
+        axes = list(range(node.inputs[0].type.ndim))
+    else:
+        axes = list(axes)
+    new_axes = [axis + batched_ndims for axis in axes]
+    new_op = op.clone(axis=new_axes)
+    return new_op.make_node(x)

pytensor/tensor/elemwise.py

@@ -5,19 +5,25 @@
 
 import pytensor
 from pytensor.configdefaults import config
-from pytensor.graph.basic import Apply, Variable
+from pytensor.graph.basic import Apply, Variable, equal_computations
 from pytensor.graph.op import Op
 from pytensor.misc.safe_asarray import _asarray
 from pytensor.scalar import ScalarVariable
 from pytensor.tensor.basic import (
     as_tensor_variable,
+    concatenate,
     constant,
     get_underlying_scalar_constant_value,
     get_vector_length,
     infer_static_shape,
 )
+from pytensor.tensor.blockwise import _vectorize_node
 from pytensor.tensor.random.type import RandomGeneratorType, RandomStateType, RandomType
-from pytensor.tensor.random.utils import normalize_size_param, params_broadcast_shapes
+from pytensor.tensor.random.utils import (
+    broadcast_params,
+    normalize_size_param,
+    params_broadcast_shapes,
+)
 from pytensor.tensor.shape import shape_tuple
 from pytensor.tensor.type import TensorType, all_dtypes
 from pytensor.tensor.type_other import NoneConst
@@ -383,3 +389,22 @@ class DefaultGeneratorMakerOp(AbstractRNGConstructor):
 
 
 default_rng = DefaultGeneratorMakerOp()
+
+
+@_vectorize_node.register(RandomVariable)
+def vectorize_random_variable(
+    op: RandomVariable, node: Apply, rng, size, dtype, *dist_params
+) -> Apply:
+    # If size was provided originally and a new size hasn't been provided,
+    # We extend it to accommodate the new input batch dimensions.
+    # Otherwise, we assume the new size already has the right values
+    old_size = node.inputs[1]
+    len_old_size = get_vector_length(old_size)
+    if len_old_size and equal_computations([old_size], [size]):
+        bcasted_param = broadcast_params(dist_params, op.ndims_params)[0]
+        new_param_ndim = (bcasted_param.type.ndim - op.ndims_params[0]) - len_old_size
+        if new_param_ndim >= 0:
+            new_size_dims = bcasted_param.shape[:new_param_ndim]
+            size = concatenate([new_size_dims, size])
+
+    return op.make_node(rng, size, dtype, *dist_params)

pytensor/tensor/random/op.py

@@ -2,6 +2,7 @@
 import pytensor.tensor.rewriting.blas
 import pytensor.tensor.rewriting.blas_c
 import pytensor.tensor.rewriting.blas_scipy
+import pytensor.tensor.rewriting.blockwise
 import pytensor.tensor.rewriting.elemwise
 import pytensor.tensor.rewriting.extra_ops
 

pytensor/tensor/rewriting/__init__.py

@@ -0,0 +1,39 @@
+from pytensor.compile.mode import optdb
+from pytensor.graph import node_rewriter
+from pytensor.graph.rewriting.basic import copy_stack_trace, out2in
+from pytensor.tensor.blockwise import Blockwise, vectorize_node
+from pytensor.tensor.rewriting.basic import register_useless
+
+
+@register_useless("fast_compile")
+@node_rewriter([Blockwise])
+def local_useless_blockwise(fgraph, node):
+    # If there is a dispatch implementation that does not require Blockwise, use that instead.
+    # This means a user created a Blockwise manually when there was no need.
+    op = node.op
+    inputs = node.inputs
+    dummy_core_node = op._create_dummy_core_node(node.inputs)
+    vect_node = vectorize_node(dummy_core_node, *inputs)
+    if not isinstance(vect_node.op, Blockwise):
+        return copy_stack_trace(node.outputs, vect_node.outputs)
+
+
+@node_rewriter([Blockwise])
+def local_useless_unbatched_blockwise(fgraph, node):
+    """Remove Blockwise that don't have any batched dims."""
+    op = node.op
+    inputs = node.inputs
+
+    if max(inp.type.ndim - len(sig) for inp, sig in zip(inputs, op.inputs_sig)) == 0:
+        return copy_stack_trace(node.outputs, op.core_op.make_node(*inputs).outputs)
+
+
+# We register this rewrite late, so that other rewrites need only target Blockwise Ops
+optdb.register(
+    "local_useless_unbatched_blockwise",
+    out2in(local_useless_unbatched_blockwise, ignore_newtrees=True),
+    "fast_run",
+    "fast_compile",
+    "blockwise",
+    position=49,
+)

pytensor/tensor/rewriting/blockwise.py

@@ -1,3 +1,5 @@
+from typing import Sequence, Union
+
 import numpy as np
 
 import pytensor
@@ -107,3 +109,54 @@ def as_list(x):
         return list(x)
     except TypeError:
         return [x]
+
+
+def import_func_from_string(func_string: str):  # -> Optional[Callable]:
+    func = getattr(np, func_string, None)
+    if func is not None:
+        return func
+
+    # Not inside NumPy or Scipy. So probably another package like scipy.
+    module = None
+    items = func_string.split(".")
+    for idx in range(1, len(items)):
+        try:
+            module = __import__(".".join(items[:idx]))
+        except ImportError:
+            break
+
+    if module:
+        for sub in items[1:]:
+            try:
+                module = getattr(module, sub)
+            except AttributeError:
+                module = None
+                break
+        return module
+
+
+def broadcast_static_dim_lengths(
+    dim_lengths: Sequence[Union[int, None]]
+) -> Union[int, None]:
+    """Apply static broadcast given static dim length of inputs (obtained from var.type.shape).
+
+    Raises
+    ------
+    ValueError
+        When static dim lengths are incompatible
+    """
+
+    dim_lengths_set = set(dim_lengths)
+    # All dim_lengths are the same
+    if len(dim_lengths_set) == 1:
+        return tuple(dim_lengths_set)[0]
+
+    # Only valid indeterminate case
+    if dim_lengths_set == {None, 1}:
+        return None
+
+    dim_lengths_set.discard(1)
+    dim_lengths_set.discard(None)
+    if len(dim_lengths_set) > 1:
+        raise ValueError
+    return tuple(dim_lengths_set)[0]

pytensor/tensor/utils.py

@@ -5,7 +5,9 @@
 from pytensor import config, function
 from pytensor.gradient import NullTypeGradError, grad
 from pytensor.raise_op import Assert
+from pytensor.tensor.blockwise import vectorize_node
 from pytensor.tensor.math import eq
+from pytensor.tensor.random import normal
 from pytensor.tensor.random.op import RandomState, RandomVariable, default_rng
 from pytensor.tensor.shape import specify_shape
 from pytensor.tensor.type import all_dtypes, iscalar, tensor
@@ -202,3 +204,37 @@ def test_RandomVariable_incompatible_size():
         ValueError, match="Size length is incompatible with batched dimensions"
     ):
         rv_op(np.zeros((2, 4, 3)), 1, size=(4,))
+
+
+def test_vectorize_node():
+    vec = tensor(shape=(None,))
+    vec.tag.test_value = [0, 0, 0]
+    mat = tensor(shape=(None, None))
+    mat.tag.test_value = [[0, 0, 0], [1, 1, 1]]
+
+    # Test without size
+    node = normal(vec).owner
+    new_inputs = node.inputs.copy()
+    new_inputs[3] = mat
+    vect_node = vectorize_node(node, *new_inputs)
+    assert vect_node.op is normal
+    assert vect_node.inputs[3] is mat
+
+    # Test with size, new size provided
+    node = normal(vec, size=(3,)).owner
+    new_inputs = node.inputs.copy()
+    new_inputs[1] = (2, 3)
+    new_inputs[3] = mat
+    vect_node = vectorize_node(node, *new_inputs)
+    assert vect_node.op is normal
+    assert tuple(vect_node.inputs[1].eval()) == (2, 3)
+    assert vect_node.inputs[3] is mat
+
+    # Test with size, new size not provided
+    node = normal(vec, size=(3,)).owner
+    new_inputs = node.inputs.copy()
+    new_inputs[3] = mat
+    vect_node = vectorize_node(node, *new_inputs)
+    assert vect_node.op is normal
+    assert vect_node.inputs[3] is mat
+    assert tuple(vect_node.inputs[1].eval({mat: mat.tag.test_value})) == (2, 3)

tests/tensor/random/test_op.py

@@ -0,0 +1,36 @@
+from pytensor import function
+from pytensor.scalar import log as scalar_log
+from pytensor.tensor import matrix, tensor3
+from pytensor.tensor.blockwise import Blockwise
+from pytensor.tensor.elemwise import Elemwise
+from pytensor.tensor.nlinalg import MatrixPinv
+
+
+def test_useless_blockwise_of_elemwise():
+    x = matrix("x")
+    out = Blockwise(Elemwise(scalar_log), signature="()->()")(x)
+
+    assert isinstance(out.owner.op, Blockwise)
+    assert isinstance(out.owner.op.core_op, Elemwise)
+
+    fn = function([x], out, mode="FAST_COMPILE")
+    assert isinstance(fn.maker.fgraph.outputs[0].owner.op, Elemwise)
+
+
+def test_useless_unbatched_blockwise():
+    x = matrix("x")
+    blockwise_op = Blockwise(MatrixPinv(hermitian=False), signature="(m,n)->(n,m)")
+    out = blockwise_op(x)
+
+    assert isinstance(out.owner.op, Blockwise)
+    assert isinstance(out.owner.op.core_op, MatrixPinv)
+
+    fn = function([x], out, mode="FAST_COMPILE")
+    assert isinstance(fn.maker.fgraph.outputs[0].owner.op, MatrixPinv)
+
+    # Test that it's not removed when there are batched dims
+    x = tensor3("x")
+    out = blockwise_op(x)
+    fn = function([x], out, mode="FAST_COMPILE")
+    assert isinstance(fn.maker.fgraph.outputs[0].owner.op, Blockwise)
+    assert isinstance(fn.maker.fgraph.outputs[0].owner.op.core_op, MatrixPinv)