NumericAL#

Dense linear algebra, tensors and the FFT whose answers are the same to the bit on any number of threads.

NumericAL is a header-only C++20 library with a Python binding that follows NumPy and SciPy: solve, eigh, svd, expm, einsum, fft and some two hundred more, on float32, float64 and their complex types. It depends on nothing but the C++ standard library; the Python package imports in about 10 ms and starts no thread until a computation pays for one.

Get started

Install the Python package or the headers, and solve your first system.

Getting started
Differences from NumPy and SciPy

What a ported program meets: defaults, refusals, dtypes, and what is not there yet.

Differences from NumPy and SciPy
Guides

Tensors and dtypes, the linear algebra, reproducibility and threads, accuracy, the BLAS ABI, the C++ library.

Guides
Python reference

Every public function and class of numerical and numerical.fft.

Python
C++ reference

numerical::matrix, tensor, the decompositions, the free functions and the exceptions.

C++
Benchmarks

Against NumPy, SciPy, OpenBLAS, Accelerate and Eigen, every answer checked first.

https://github.com/RECHE23/NumericAL/blob/main/BENCHMARKS.md

What it guarantees#

  • The same bits on one thread as on all, and on every x86-64 processor with AVX2 or AVX-512 for one build: each sum runs in one written order, whatever the threads or the vector width.

  • A refusal rather than a meaningless answer. A NaN or an infinity in a matrix raises LinAlgError; a singular system raises; an ill-conditioned one warns.

  • Backward-stable answers within stated bounds, held by differential fuzzers against NumPy and SciPy on adversarial matrices (see Accuracy).

  • No dependency and no idle spinning: no BLAS underneath, no thread busy-waiting between calls, safe after fork.

A first look#

>>> import numerical
>>> a = numerical.asarray([[2.0, 0.0], [0.0, 4.0]])
>>> numerical.solve(a, [2.0, 2.0])
numerical.Tensor([1.0, 0.5], dtype=float64)
>>> numerical.einsum("ii->", a)
numerical.Tensor(6.0, dtype=float64)
>>> numerical.vector_norm([1e200, 1e200])
1.414213562373095e+200