Getting started#

Python#

pip install numeric-al        # the distribution is numeric-al; the module is numerical

The wheels serve CPython 3.10 and later through the stable ABI, and the free-threaded 3.14t and 3.15t without the GIL, on Linux (x86-64, arm64), macOS (arm64, x86-64) and Windows (x86-64). Every function takes a numerical.Tensor or anything numerical.asarray() reads: a NumPy array, a nested list, a number.

>>> import numerical
>>> a = [[4.0, 1.0], [1.0, 3.0]]
>>> numerical.solve(a, [1.0, 2.0])
numerical.Tensor([0.09090909090909091, 0.6363636363636364], dtype=float64)
>>> values, vectors = numerical.eigh(a)
>>> numerical.round(values, 12)
numerical.Tensor([2.38196601125, 4.61803398875], dtype=float64)
>>> t = numerical.arange(6).reshape(2, 3)
>>> t.sum(axis=0)
numerical.Tensor([3.0, 5.0, 7.0], dtype=float64)
>>> numerical.fft.rfft([1.0, 2.0, 3.0, 4.0])
numerical.Tensor([(10+0j), (-2+2j), (-2+0j)], dtype=complex128)

A Tensor passes to NumPy, PyTorch, JAX or CuPy without a copy (np.asarray(t), np.from_dlpack(t), torch.from_dlpack(t)); numerical.asarray(array) and numerical.from_dlpack(x) (a PyTorch or JAX CPU tensor) copy the other way. Tensors and dtypes says what a Tensor holds and how the dtypes promote; Linear algebra lists the linear algebra, and Differences from NumPy and SciPy what a program written for NumPy or SciPy meets. numpy.linalg’s names are also in numerical.linalg, and numpy.fft’s (fft, ifft, fftn, rfft, irfft, fftfreq…) and scipy.fft’s cosine and sine transforms (dct, idct, dst, idst, dctn…) in numerical.fft.

C++#

Header-only: add include/ to the include path, or after cmake --install, in your CMakeLists.txt:

find_package(numerical CONFIG REQUIRED)
target_link_libraries(your_target PRIVATE numerical::numerical) # C++20, and the threads library

then

#include <cstdio>

#include <numerical/numerical.hpp>

int main()
{
  const numerical::matrix<double> a {{4.0, 1.0, 0.0}, {1.0, 3.0, 1.0}, {0.0, 1.0, 2.0}};
  const numerical::vector<double> b {1.0, 2.0, 3.0};

  const numerical::vector<double>  x     {numerical::solve(a, b)}; // LU, partial pivoting
  const numerical::symmetric_eigen eigen {a};                      // A = V L V^T, as numpy.linalg.eigh
  std::printf("x = (%g, %g, %g), smallest eigenvalue %g\n", x[0], x[1], x[2], eigen.values()[0]);
}

numerical.get_include() gives the header directory of an installed Python package, so a C++ extension can build against the same headers. With GCC, compile with -ffp-contract=on (the wheels’ setting), and with Clang for arm64 with -mllvm -force-ordered-reductions=false; Reproducibility and threads says why. The C++ library lists the classes, the supported toolchains and how to build from source.