CoreMLMath

Multithreaded C++ backend for Machine Learning Math.

Thanks to pybind11 and OpenMP for existing in this world.

Financial Support

Build Instructions

• Clone the repo locally:

  git clone https://github.com/novfensec/coremlmath -b master --single-branch --depth 1
  cd coremlmath
  

• Linux:

  sudo apt update
  sudo apt install -y build-essential cmake
  

• Windows (Powershell):

  Install the necessary C++ build tools and CMake using Windows Package Manager (winget):

  This has to be done only once.

  winget install -e --id Microsoft.VisualStudio.BuildTools --override "--passive --wait --add Microsoft.VisualStudio.Workload.VCTools --includeRecommended" --source winget
  winget install -e --id Kitware.CMake --source winget
  

• Install pybind11:

  pip install "pybind11[global]" uv
  

• Build and install using standard pip:

  pip install .
  

  Or using uv for isolated, fast builds:

  uv pip install .
  

• Build an installable wheel:

  uv build --wheel
  

Usage

The package expects 2D and 1D NumPy arrays of integers for maximum memory efficiency.

import numpy as np
import cmlmath
import time

print("Generating massive dataset (1,000,000 rows, 500 features)...")
X_raw = np.random.randint(0, 2, size=(1000000, 500), dtype=np.int32)
y_raw = np.random.randint(0, 16, size=(1000000,), dtype=np.int32)

# Enforce C-contiguous (Row-Major) layout in memory. 
# (np.random generates C-contiguous arrays by default, but this guarantees 
# safety if you ever swap the data source out for a Pandas DataFrame or transposed array).
X = np.ascontiguousarray(X_raw, dtype=np.int32)
y = np.ascontiguousarray(y_raw, dtype=np.int32)

print(f"Starting multithreaded C++ calculation using {cmlmath.get_cpu_cores()} cores...")
start = time.time()

info_gains = cmlmath.information_gain(X, y)
gini_gains = cmlmath.gini_gain(X, y)

calc_time = time.time() - start

print(f"Calculated 500 Information and Gini Gains in {calc_time:.4f} seconds.")
print(f"Best feature index to split on (Information Gain): {np.argmax(info_gains)}")
print(f"Best feature index to split on (Gini Gain): {np.argmax(gini_gains)}")