gpu-prime-generator

GPU Prime Number Generator

This project provides a CUDA-accelerated Python script that generates an ordered list of prime numbers using an NVIDIA GPU via Numba. It streams primes to a text file and supports safe stop/restart.

Features

• CUDA-accelerated prime checking using a segmented sieve approach.
• Ordered output: primes are written in ascending order.
• Graceful shutdown: press Ctrl+C to stop; the script saves the current state.
• Resume capability: on restart, the script continues generating from the last prime.
• Configurable segment size: adjust how many numbers are tested per GPU batch.

Prerequisites

1. Python 3.9+

2. CUDA Toolkit (Runtime + NVVM) or conda cuda package

   • Recommended:

     conda create -n primegpu python=3.11 numba cuda -c nvidia -c conda-forge
     conda activate primegpu
     pip install numpy

   • Alternatively, install the CUDA Toolkit from NVIDIA (ensure nvvm.dll is on your PATH).

3. Numba (0.59+ for Python 3.12; 0.61+ for Python 3.13)

   pip install numba

4. NumPy

   pip install numpy

Verify that Numba sees your GPU:

from numba import cuda
print(cuda.detect())

Installation

1. Clone this repository or download the code:

   git clone https://github.com/yourusername/gpu-prime-generator
   cd gpu-prime-generator

2. Ensure your Python environment has the prerequisites (see above).

Usage

Run the script with default settings:

python gpu_prime_generator.py

• Creates (or appends to) primes.txt in the current folder.
• Press Ctrl+C to stop. The script prints a summary, for example:

  Stopped by user. Found  784,984 primes in 12.4 s  (63 k p/s)
  

Options

• -o, --output <file>
  Destination file (default: primes.txt). If the file exists, the script resumes from the last prime.

• -n, --segment <count>
  How many integers to test per GPU batch (default: 1000000).

Example:

python gpu_prime_generator.py -o my_primes.txt -n 5000000

How It Works

1. CPU sieve (bootstrap)

   • Computes all primes up to √N for each segment using a standard Sieve of Eratosthenes.

2. GPU kernel

   • Each CUDA thread tests one odd candidate number by checking divisibility against the pre-sieved primes.
   • Results are streamed back, filtered, and written to the output file in ascending order.

3. Segmented loop

   • Numbers are processed in chunks (segments) of configurable size to manage memory usage.
   • After each segment, the script flushes new primes to disk and prints progress.

4. Graceful shutdown / resume

   • On startup, the script reads the last prime in the output file (if any) and keeps a running total.
   • Pressing Ctrl+C triggers a KeyboardInterrupt, closing the file and printing a final summary.

Tips & Tuning

• Increase --segment for fewer kernel launches (uses more GPU memory).
• Adjust threads_per_block in the source code (128–512 threads is typical).
• For best write performance, use an SSD or fast NVMe drive.
• Ensure only one CUDA installation is on your PATH to avoid conflicts.

Troubleshooting

• nvvm.dll not found
  Ensure the CUDA Toolkit (runtime + NVVM) is installed and its bin directories are on PATH.
  For conda users, install the cuda meta-package:

  conda install -c nvidia -c conda-forge cuda

• CUDA driver mismatch
  Make sure your GPU driver version is ≥ the CUDA Toolkit version you installed.

• Python / Numba compatibility

  • Python 3.12 requires Numba 0.59+
  • Python 3.13 requires Numba 0.61+

License

This project is released under the MIT License. See LICENSE for details.