linac 1.0.0

Linear Algebra with CUDA

linac_logo_v1

Linac implements GPU-accelerated linear algebra using CUDA, with support for real and complex floating-point arithmetic, finite fields, and leading-digit p-adic numbers. Its main functionality is dense Gaussian elimination to reduced row-echelon form. Applications include functional reconstruction of analytic scattering amplitudes.

Installation

The latest stable release is available from PyPI:

pip install linac

To enable GPU acceleration, install the CUDA extra:

pip install "linac[cuda]"

For development, clone the repository and install it in editable mode:

git clone https://github.com/GDeLaurentis/linac.git
pip install -e "linac[full]"

Available extras are:

cuda, dev, full(=cuda+dev)

Requirements

The core package depends on:

numpy, pycoretools, pyadic, syngular

Optional dependencies include:

pycuda   # CUDA/GPU support

with the cuda extra, and

mpmath, galois, diskcache, pytest, pytest-cov, flake8

with the dev extra.

GPU acceleration requires a working CUDA development environment, including nvcc, compatible NVIDIA drivers, and access to a CUDA-enabled GPU.

You can check your CUDA setup with:

nvcc --version
nvidia-smi

Quick Start

The main entry point is cuda_row_reduce, which computes a row-reduced echelon form on the GPU.

import numpy
from linac import cuda_row_reduce

p = 2**31 - 1

A = numpy.random.randint(0, p, size=(2000, 2000), dtype="uint32")
rref = cuda_row_reduce(A, field_characteristic=p)

Setting field_characteristic=0 selects floating-point arithmetic:

A = numpy.random.rand(2000, 2000)
rref = cuda_row_reduce(A, field_characteristic=0)

For CPU row reduction, use:

from linac import row_reduce

rref, variable_order = row_reduce(A, prime=p)

Linac also provides utilities for constructing dense linear systems from polynomial ansätze:

from linac import load_matrices

matrices = load_matrices(prefactors, ansatze, points, use_cuda=True)

Testing

Run the test suite with:

pytest -rs --verbose --cov=linac --cov-report=html

Timings

The following timings are for dense Gaussian elimination to reduced row-echelon form over the finite field with p=(2**31 - 1). Times are shown in seconds as mean ± standard deviation.

Matrix size	A100 80GB	RTX 4070 Laptop 8GB
1,000	0.52 ± 0.01	0.28 ± 0.02
2,000	0.56 ± 0.01	0.34 ± 0.01
5,000	1.02 ± 0.01	3.31 ± 0.01
10,000	3.60 ± 0.01	25.97 ± 0.66
15,000	10.23 ± 0.02	98.21 ± 0.85
20,000	23.40 ± 0.03	241.29 ± 3.00
100,000	2824.64	—

The 100k run used approximately 37.67 GiB of memory on an 80 GiB A100.

For full benchmark details and comparison with CPU implementations, see the accompanying paper and documentation.

Size Limit

For 32-bit entries, the approximate upper bound for a dense square matrix is

N_max ~ sqrt(VRAM [bytes] / 4)

Available VRAM	Approx. square matrix size
4 GB	31,622
8 GB	44,721
11 GB	52,440
12 GB	54,772
16 GB	63,245
24 GB	77,459
40 GB	100,000
80 GB	141,421

The practical limit is usually close to this estimate, up to padding and (small) memory-management overheads.

Documentation

Documentation is available at:

https://gdelaurentis.github.io/linac/

Citation

If you use Linac in academic work, please cite the accompanying paper and the Zenodo archive:

@software{linac,
  author = {De Laurentis, Giuseppe and Franklin, Jack},
  title = {Linac: linear algebra with CUDA over finite fields},
  year = {2026},
  doi = {10.5281/zenodo.xxxxxxxx},
}