Awesome

numpy-minmax: a fast function for finding the minimum and maximum value in a NumPy array

NumPy lacked an optimized minmax function, so we wrote our own. At Nomono, we use it for audio processing, but it can be applied any kind of float32 ndarray.

• Written in C and takes advantage of AVX/AVX512 for speed
• Roughly 2.3x speedup compared to the numpy amin+amax equivalent (tested on Intel CPU with numpy 1.24-1.26)
• The fast implementation is tailored for float32 arrays that are C-contiguous, F-contiguous or 1D strided. Strided arrays with ndim >= 2 get processed with numpy.amin and numpy.amax, so no perf gain there. There is also a fast implementation for contiguous int16 arrays.

Installation

$ pip install numpy-minmax

Usage

import numpy_minmax
import numpy as np

arr = np.arange(1337, dtype=np.float32)
min_val, max_val = numpy_minmax.minmax(arr)  # 0.0, 1336.0

Changelog

[0.3.1] - 2024-08-15

Changes

• Optimize (with AVX) the processing of contiguous int16 arrays. ~2.3x speedup compared to 0.3.0

For the complete changelog, go to CHANGELOG.md

Development

• Install dev/build/test dependencies as denoted in pyproject.toml
• CC=clang pip install -e .
• pytest

Running benchmarks

• Install diplib pip install diplib
• python scripts/perf_benchmark.py

Acknowledgements

This library is maintained/backed by Nomono, a Norwegian audio AI startup.