NumPy ufunc Introduction

What Is a ufunc?

A universal function (ufunc) is a function that operates on an ndarray in an element-by-element fashion, supporting broadcasting, type casting, and several other standard features. Functions like np.add, np.sin, and np.sqrt are all ufuncs. They are implemented in compiled C code, which is why they run dramatically faster than an equivalent Python for loop.

Vectorization: Why ufuncs Matter

Without ufuncs, adding two lists element-wise in plain Python requires a loop or a list comprehension that iterates one item at a time and pays the cost of Python's interpreter overhead on every step. A ufunc pushes that entire loop down into optimized, pre-compiled C code, so the same operation on millions of elements can be an order of magnitude faster.

import numpy as np

a = [1, 2, 3, 4]
b = [10, 20, 30, 40]

# Without a ufunc -- a manual Python loop
result = []
for i in range(len(a)):
    result.append(a[i] + b[i])
print(result)

# With a ufunc -- vectorized and much faster on large arrays
result = np.add(a, b)
print(result)

Checking Whether a Function Is a ufunc

You can confirm a function is a ufunc by inspecting its type with the built-in type() function. Genuine ufuncs report their type as numpy.ufunc, which distinguishes them from regular Python functions defined in the numpy namespace.

import numpy as np

print(type(np.add))        # <class 'numpy.ufunc'>
print(type(np.multiply))   # <class 'numpy.ufunc'>

def my_add(a, b):
    return a + b

print(type(my_add))        # <class 'function'>

• They operate element-wise, so the output array has the same shape as the (broadcast) input.
• They support broadcasting between arrays of different but compatible shapes.
• They can accept an optional out parameter to write results into an existing array without allocating new memory.
• Many ufuncs support a where parameter to selectively apply the operation.
• They are implemented in C, so they avoid Python-level loop overhead entirely.

import numpy as np

a = np.array([1, 2, 3, 4, 5])
b = np.array([10, 20, 30, 40, 50])

# Write results directly into a pre-allocated array
result = np.zeros(5)
np.add(a, b, out=result)
print(result)

# Broadcasting: a smaller array is stretched to match a larger one
scaled = np.multiply(a, 10)
print(scaled)