Avoid numpy broadcast_to and ndindex in hot loops

Author: ricardoV94

pytensor/tensor/random/basic.py

@@ -18,6 +18,7 @@
     broadcast_params,
     normalize_size_param,
 )
+from pytensor.tensor.utils import faster_broadcast_to, faster_ndindex
 
 
 # Scipy.stats is considerably slow to import
@@ -976,19 +977,13 @@ def __call__(self, alphas, size=None, **kwargs):
     @classmethod
     def rng_fn(cls, rng, alphas, size):
         if alphas.ndim > 1:
-            if size is None:
-                size = ()
-
-            size = tuple(np.atleast_1d(size))
-
-            if size:
-                alphas = np.broadcast_to(alphas, size + alphas.shape[-1:])
+            if size is not None:
+                alphas = faster_broadcast_to(alphas, size + alphas.shape[-1:])
 
             samples_shape = alphas.shape
             samples = np.empty(samples_shape)
-            for index in np.ndindex(*samples_shape[:-1]):
+            for index in faster_ndindex(*samples_shape[:-1]):
                 samples[index] = rng.dirichlet(alphas[index])
-
             return samples
         else:
             return rng.dirichlet(alphas, size=size)
@@ -1800,11 +1795,11 @@ def rng_fn(cls, rng, n, p, size):
             if size is None:
                 n, p = broadcast_params([n, p], [0, 1])
             else:
-                n = np.broadcast_to(n, size)
-                p = np.broadcast_to(p, size + p.shape[-1:])
+                n = faster_broadcast_to(n, size)
+                p = faster_broadcast_to(p, size + p.shape[-1:])
 
             res = np.empty(p.shape, dtype=cls.dtype)
-            for idx in np.ndindex(p.shape[:-1]):
+            for idx in faster_ndindex(p.shape[:-1]):
                 res[idx] = rng.multinomial(n[idx], p[idx])
             return res
         else:
@@ -1978,13 +1973,13 @@ def rng_fn(self, *params):
                     p.shape[:batch_ndim],
                 )
 
-        a = np.broadcast_to(a, size + a.shape[batch_ndim:])
+        a = faster_broadcast_to(a, size + a.shape[batch_ndim:])
         if p is not None:
-            p = np.broadcast_to(p, size + p.shape[batch_ndim:])
+            p = faster_broadcast_to(p, size + p.shape[batch_ndim:])
 
         a_indexed_shape = a.shape[len(size) + 1 :]
         out = np.empty(size + core_shape + a_indexed_shape, dtype=a.dtype)
-        for idx in np.ndindex(size):
+        for idx in faster_ndindex(size):
             out[idx] = rng.choice(
                 a[idx], p=None if p is None else p[idx], size=core_shape, replace=False
             )
@@ -2097,10 +2092,10 @@ def rng_fn(self, rng, x, size):
             if size is None:
                 size = x.shape[:batch_ndim]
             else:
-                x = np.broadcast_to(x, size + x.shape[batch_ndim:])
+                x = faster_broadcast_to(x, size + x.shape[batch_ndim:])
 
             out = np.empty(size + x.shape[batch_ndim:], dtype=x.dtype)
-            for idx in np.ndindex(size):
+            for idx in faster_ndindex(size):
                 out[idx] = rng.permutation(x[idx])
             return out
 

pytensor/tensor/random/utils.py

@@ -15,6 +15,7 @@
 from pytensor.tensor.math import maximum
 from pytensor.tensor.shape import shape_padleft, specify_shape
 from pytensor.tensor.type import int_dtypes
+from pytensor.tensor.utils import faster_broadcast_to
 from pytensor.tensor.variable import TensorVariable
 
 
@@ -125,7 +126,7 @@ def broadcast_params(
     shapes = params_broadcast_shapes(
         param_shapes, ndims_params, use_pytensor=use_pytensor
     )
-    broadcast_to_fn = broadcast_to if use_pytensor else np.broadcast_to
+    broadcast_to_fn = broadcast_to if use_pytensor else faster_broadcast_to
 
     # strict=False because we are in a hot loop
     bcast_params = [

pytensor/tensor/utils.py

@@ -1,8 +1,10 @@
 import re
 from collections.abc import Sequence
+from itertools import product
 from typing import cast
 
 import numpy as np
+from numpy import nditer
 
 import pytensor
 from pytensor.graph import FunctionGraph, Variable
@@ -233,3 +235,22 @@ def normalize_reduce_axis(axis, ndim: int) -> tuple[int, ...] | None:
 
     # TODO: If axis tuple is equivalent to None, return None for more canonicalization?
     return cast(tuple, axis)
+
+
+def faster_broadcast_to(x, shape):
+    # Stripped down core logic of `np.broadcast_to`
+    return nditer(
+        (x,),
+        flags=["multi_index", "zerosize_ok"],
+        op_flags=["readonly"],
+        itershape=shape,
+        order="C",
+    ).itviews[0]
+
+
+def faster_ndindex(*shape):
+    """Equivalent to `np.ndindex` but usually 10x faster.
+
+    https://github.com/numpy/numpy/issues/28921
+    """
+    return product(*(range(s) for s in shape))