Performance Tips: Vectorization, Broadcasting, and Memory Layout

Performance Tips for NumPy

Writing fast NumPy code is about thinking in whole‑array operations.

Vectorize instead of looping

# Slow Python loop
s = 0
for i in range(1_000_000):
    s += i*i

# Fast NumPy vectorized version
arr = np.arange(1_000_000)
s_fast = np.sum(arr**2)

Use in‑place operations to reduce temporary arrays

arr = np.arange(10, dtype=np.float64)
arr *= 2   # modifies arr in place, no new array created

Control memory layout using order

C_contig = np.arange(12).reshape(3,4, order='C')   # row‑major
F_contig = np.arange(12).reshape(3,4, order='F')   # column‑major

Avoid unnecessary copies with np.r_ and np.c_

# Concatenation without extra copy
a = np.arange(3)
b = np.arange(3,6)
combined = np.r_[a, b]

Use np.einsum for complex summations

# Compute trace of a matrix efficiently
M = np.random.rand(5,5)
trace = np.einsum('ii', M)