cuda-kernels 0.2.0

CUDA accelerated correlation and sum reduction functions

CUDA Kernels

A Python package providing CUDA-accelerated functions for autocorrelation and sum reduction operations, with automatic CPU fallback when CUDA is not available.

Installation

From PyPI (Recommended)

pip install cuda-kernels

The package is distributed as a source distribution and compiled on your machine at install time. If the CUDA toolkit (nvcc) is detected, the GPU kernels are built and used automatically; otherwise the package installs CPU-only and runs the optimized NumPy fallback. GPU acceleration therefore requires nvcc (CUDA Toolkit) and a C/C++ host compiler to be present when you run pip install.

From GitHub

pip install git+https://github.com/AstuteFern/cuda-toolkit.git

From Source

git clone https://github.com/AstuteFern/cuda-toolkit.git
cd cuda-toolkit
pip install .

Requirements

• Python 3.6+
• NumPy

Optional (for CUDA acceleration)

• NVIDIA GPU with CUDA support
• CUDA Toolkit (version 11.0+)

Note: The package works on any system. If CUDA is not available, it automatically uses optimized CPU implementations.

Quick Start

import numpy as np
from cuda_kernels import autocorrelation, reduction_sum

# Create test data
data = np.random.randn(1000).astype(np.float32)

# Compute autocorrelation (automatically uses CUDA if available)
acf = autocorrelation(data, max_lag=50)
print(f"Autocorrelation shape: {acf.shape}")

# Compute sum reduction
total = reduction_sum(data)
print(f"Sum: {total}")

API Reference

autocorrelation(data, max_lag=None, force_cpu=False)

Compute autocorrelation of a time series.

Parameters:

• data (numpy.ndarray): Input 1D array (converted to float32)
• max_lag (int, optional): Maximum lag to compute. Default: len(data)-1
• force_cpu (bool): Force CPU implementation. Default: False

Returns:

• numpy.ndarray: Autocorrelation values for lags [0, max_lag)

reduction_sum(data, force_cpu=False)

Compute sum of array elements.

Parameters:

• data (numpy.ndarray): Input 1D array (converted to float32)
• force_cpu (bool): Force CPU implementation. Default: False

Returns:

• float: Sum of all elements

Examples

Basic Usage

import numpy as np
from cuda_kernels import autocorrelation, reduction_sum

# Example 1: Autocorrelation
signal = np.sin(np.linspace(0, 4*np.pi, 1000)).astype(np.float32)
acf = autocorrelation(signal, max_lag=100)

# Example 2: Sum reduction
data = np.array([1, 2, 3, 4, 5], dtype=np.float32)
total = reduction_sum(data)  # Returns 15.0

Checking CUDA Status

import sys
autocorr_module = sys.modules['cuda_kernels.autocorrelation']
reduction_module = sys.modules['cuda_kernels.reduction']

print(f"CUDA available: {autocorr_module._cuda_available}")

Force CPU Mode

# Useful for testing or when you want consistent behavior
cpu_result = reduction_sum(data, force_cpu=True)

Performance

• With CUDA: Significant speedup for large arrays (10K+ elements)
• CPU Fallback: Optimized NumPy implementations, still efficient for most use cases
• Automatic Detection: No configuration needed, works out of the box

License

MIT License - see LICENSE file for details.