Minimal Numpy: Three Ways to Build a Vector Library in Python

This project demonstrates how to implement a minimal vector library in Python using three different native extension methods:

• by_ctypes: C library loaded via Python's ctypes
• by_python_c_api: C extension using the Python C API
• by_pybind11: C++ library exposed via pybind11

Each approach provides a Vector class supporting element-wise addition and subtraction.

---

Structure

by_ctypes/
	minimal_numpy.c      # C implementation of vector operations
	minimal_numpy.py     # Python wrapper using ctypes
	Makefile             # Build shared library

by_python_c_api/
	minimal_numpy.c      # C extension using Python C API
	setup.py             # Build script
	test.py              # Usage example
	README.md            # Approach-specific instructions

by_pybind11/
	minimal_numpy.cpp    # C++ implementation with pybind11 bindings
	setup.py             # Build script
	test.py              # Usage example
	README.md            # Approach-specific instructions

---

Quick Start

by_ctypes

Build:

cd by_ctypes
make run

by_python_c_api and by_pybind11

See their respective README.md files for build and run instructions.

---

Features

• Vector class: stores a list of doubles
• Supports + and - operators for element-wise addition/subtraction
• Type and size checks for safe operations
• Demonstrates three native extension techniques
• Checked for AVX instruction usage in compiled libraries

Performance

We run size = 1000000 and repetition = 50 for addition and subtraction benchmarks. 10 warm-up runs are performed before timing to ensure fair measurements. Here is a summary of the results (lower is better):

Method	Addition Time (s)	Subtraction Time (s)
by_ctypes	2.4886	2.4444
by_python_c_api	0.0171	0.0170
by_pybind11	0.0392	0.0389
numpy_reference	0.0203	0.0208

Development Environment

For clangd user, just run pkg-config --cflags python3 to get the necessary include paths and add them to .clangd: