numpy-low-level

NumPy - Low-Level Optimization & Memory

At high volumes, standard NumPy operations can still be slow due to unnecessary memory allocations. This guide covers how to manipulate the internal representation of arrays to achieve C-level performance without leaving Python.

When to Use

• Implementing sliding window algorithms (convolutions) without extra memory.
• Interfacing Python with C, C++, or Fortran code via pointers.
• Working with complex, heterogeneous data structures (Structured Arrays).
• Optimizing memory-constrained systems via Memory Mapping (memmap).
• Debugging performance issues related to "Memory Layout" (C-style vs Fortran-style).

Core Principles

1. The Metadata vs. Data Split

A NumPy array is a small Header (shape, dtype, strides) pointing to a large Data Buffer. Many operations (like .T, reshape, slice) only change the Header. This is "Zero-Copy".

2. Strides (The Step Logic)

Strides define how many bytes to skip in memory to get to the next element in each dimension. Manipulating strides allows you to "cheat" and create virtual views of data.

3. Contiguity

• C-Contiguous: Last index varies fastest (Row-major).
• F-Contiguous: First index varies fastest (Column-major).
• Vectorization is significantly faster on contiguous memory.

Quick Reference: Memory Inspection

import numpy as np

arr = np.zeros((100, 100))

print(arr.flags)         # Check contiguity and ownership
print(arr.strides)       # bytes to step in each axis
print(arr.__array_interface__['data']) # Memory pointer address

Critical Rules

✅ DO

• Prefer Views over Copies - Use slicing and reshaping whenever possible.
• Check base - Use arr.base is None to verify if an array owns its memory or is just a view.
• Use Structured Arrays - For "Table of Records" data where you need NumPy speed but different types per column.
• Align Memory - Ensure arrays are aligned to 64-bit boundaries for SIMD optimization.
• Use out= parameters - Most NumPy functions accept an out argument to prevent creating a new temporary array.