NumericAL

A pure, header-only C++20 numerical computing library — a leaf of the SciLang ecosystem (a sibling of REAL), generic over its scalar type and depending on nothing but the C++ standard library.

NumericAL is agnostic of expressions and of any symbolic layer: it computes with concrete numeric structures — vectors, matrices, tensors — and the algorithms over them. SciLang consumes it as a numeric backend through its module seam; the symbolic↔numeric reconciliation stays in SciLang.

The API reference, generated from the headers: https://reche23.github.io/NumericAL/.

Quick start (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 (this is cmake/smoke/main.cpp, which CI builds against the installed package):

#include <cmath>
#include <complex>
#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
  const numerical::eigen_decomposition<double> general {a};                      // as numpy.linalg.eig
  const numerical::svd_decomposition<double>   svd     {a};                      // A = U S V^H

  std::printf("x = %.6f %.6f %.6f\n", x[0], x[1], x[2]);
  std::printf("eigh: %.6f %.6f %.6f\n", eigen.values()[0], eigen.values()[1], eigen.values()[2]);
  std::printf("eig: %.6f%+.6fi, its vectors %zu x %zu\n", general.values()[0].real(), general.values()[0].imag(),
              general.vectors().rows(), general.vectors().cols());
  std::printf("svd: %.6f %.6f %.6f\n", svd.singular_values()[0], svd.singular_values()[1], svd.singular_values()[2]);
  const numerical::vector<double> residual {(a * x) - b};
  return std::abs(residual[0]) + std::abs(residual[1]) + std::abs(residual[2]) < 1e-12 ? 0 : 1;
}

The Python package, pip install numeric-al, follows NumPy's API (see Python binding, below).

Capabilities

Small and complete, grown measured, honest about what is not yet here:

• numerical::vector<T> — a dense, value-semantic vector: element-wise arithmetic, scalar multiplication, dot product, tolerance comparison, shape checks.

• numerical::matrix<T> — a dense, row-major, value-semantic matrix: element-wise arithmetic, scalar / matrix–vector / matrix–matrix products, transpose, tolerance comparison, shape checks.

• numerical::lu_decomposition<T> — an LU factorisation (partial pivoting) computed once and reused: solve (any number of right-hand sides), determinant, and inverse. Free solve / determinant / inverse are one-shot conveniences.

• numerical::cholesky_decomposition<T> — A = L Lᴴ for a Hermitian positive-definite matrix (about twice as cheap as LU): the factor lower(), solve, and determinant; throws not_positive_definite_error otherwise.

• numerical::qr_decomposition<T> — A = Q R (Householder reflections, so Q stays orthonormal to rounding however ill-conditioned A is) for an m × n matrix of any shape, as NumPy's reduced qr: the factors q() (m × k) / r() (k × n), k = min(m, n) (qr_mode::complete: q() m × m, r() m × n, the reduced factors extended), and for m ≥ n a least-squares solve of A x ≈ b, which throws singular_matrix_error when A is rank-deficient.

• numerical::hermitian_eigen — the same for a complex Hermitian matrix (real eigenvalues, a unitary V), reduced to a real tridiagonal by Householder reflections and unit phases, then diagonalised by divide and conquer, as LAPACK's zheevd does.

• numerical::symmetric_eigen — eigenvalues (ascending) and eigenvectors of a real symmetric matrix (A = V Λ Vᵀ): Householder tridiagonalisation by panels then divide and conquer, as LAPACK's dsyevd (most of the work in matrix products); implicit QL (eigen_method::tridiagonal_ql, dsyev's) or cyclic Jacobi rotations (eigen_method::jacobi) on request.

• symmetric_eigen(a, b), hermitian_eigen(a, b) — the pencil A x = λ B x, B positive definite, as scipy.linalg.eigh(a, b) (LAPACK's sygvd): B = L Lᴴ, L⁻¹ A L⁻ᴴ diagonalised, the vectors B-orthonormal.

• numerical::eigen_decomposition<T> — the eigenvalues and right eigenvectors of a general real or complex square matrix, complex either way, as numpy.linalg.eig (LAPACK's dgeev/zgeev: balancing, blocked Hessenberg reduction, the multishift QR algorithm with aggressive early deflation); values(), vectors(), and eigenvalues(a) without the vectors (as eigvals).

• numerical::schur_decomposition<T> — A = Z T Zᴴ, as scipy.linalg.schur (LAPACK's dgees/zgees, from the same reduction and iterations): t() quasi-triangular for a real matrix (a pair a 2 × 2 block in standard form), triangular for a complex one, z() orthogonal or unitary.

• numerical::svd_decomposition<T> — the singular value decomposition A = U Σ Vᴴ of a real or complex m × n matrix, any shape: a Householder QR, then Golub–Kahan bidiagonalisation of R (by panels, for a real one) and divide and conquer, as LAPACK's dgesdd; implicit-shift QR (svd_method::bidiagonal_qr, dgesvd's) or one-sided Jacobi on R (svd_method::jacobi, which gives every singular value to high relative accuracy) on request. singular_values() (descending), u() (m × k), v() (n × k) and vh(), with k = min(m, n) (full = true: u() m × m and v() n × n, the thin factors extended by an orthonormal basis of the complement); U and V stay orthonormal for a rank-deficient A.

• numerical::solve_triangular(a, b, lower, op, unit_diagonal) — A X = B, Aᵀ X = B or Aᴴ X = B (triangular_op) for a triangular A, as SciPy's solve_triangular: its lower or upper triangle read, the other never; n² k operations, where solve factorises first.

• numerical::lstsq(a, b) — the minimum-norm least-squares solution of A x ≈ b, as NumPy's lstsq gives it, for any shape and rank (from the SVD), b a vector or a matrix of right-hand sides (as solve(a, b) takes too); least_squares(a, b) adds the residuals, the rank and the singular values NumPy returns beside it; pinv(a) from the same SVD, singular_values(a) without the singular vectors (by dqds, as LAPACK's dlasq1: a bidiagonal's each to high relative accuracy; and matrix_rank(a) from them), symmetric_eigenvalues(a) without the eigenvectors (as NumPy's eigvalsh), and lu_decomposition::log_determinant() — NumPy's slogdet, which does not underflow where the determinant does. lstsq, least_squares, pinv and matrix_rank take a rank_cutoff: NumPy's default, rank_cutoff::relative(r) (rcond, rtol) or rank_cutoff::absolute(t) (tol); pinv(a, cutoff, true) and matrix_rank(a, cutoff, true) take a Hermitian a by its eigendecomposition, its lower triangle alone read.

• numerical::tensor<T> — a dense, row-major, N-dimensional tensor (rank-1 is a vector, rank-2 a matrix, rank-0 a scalar; built directly or from a vector / matrix), with einsum — Einstein-summation index notation for contractions and general products, as in NumPy / PyTorch. einsum is the unifying primitive of the algebra: "ij,jk->ik" is the matrix product, "i,i->" the dot product, "ij->ji" the transpose, "ii->" the trace, "i,j->ij" the outer product, "ij->i" a row reduction. It takes one or more operands and sums over every label absent from the output, in a deterministic order.

• numerical::fft (fft.hpp) — the discrete Fourier transform along an axis, as numpy.fft: fft, ifft, rfft, irfft, hfft, ihfft and fftshift, NumPy's points (cut or padded with zeros) and scalings (fft_norm). Any length: Stockham passes of radices 2 to 16 and every prime to 61, Bluestein's convolution past them; twiddle factors computed in double-double, so each is the correctly rounded root. Two sets of bits only, each the same on every machine and compiler: the portable path's (products rounded apart) and the vector paths' (NEON, AMX, AVX2 and AVX-512: one fused complex product).

• dct, idct, dst, idst (fft.hpp) — the cosine and sine transforms of scipy.fft, types 1 to 4, its norms and orthogonalize, by the real (types 1 to 3) or complex (type 4) transform: DCT-II by Makhoul's reordering, DCT-III its inverse, the sines from the cosines.

• Arithmetic on tensors (tensor_ops.hpp): + - * / element by element with NumPy's broadcasting, sum, mean, max, min of every element or along an axis (sums by NumPy's pairwise blocks), reshape, and zeros, ones, full, eye, arange, linspace.

• numerical::blas (blas.hpp) — dot, gemm (of float, double and their complex numbers), syrk, gemv, trsm, gesv, getrf, getrs, potrf and potrs (the last five, as gemm, of float, double and their complex numbers), syevd, gesdd, geqrf, orgqr, gelsd and geev with the column-major semantics of BLAS's ddot, dgemm, dsyrk, dgemv and dtrsm and LAPACK's routines of those names; the numerical_blas shared library exports them as dgemm_, dgesv_… (Fortran calling convention, 32-bit integers), to stand in for those routines under a program linked against BLAS and LAPACK (LD_PRELOAD) or a switch such as FlexiBLAS. The Linux wheels carry it (see BLAS and LAPACK for other programs); CMake builds it with -DNUMERICAL_BUILD_BLAS=ON.

Grows in as measured: sparse storage, banded and LDLᵀ factorisations, the rest of the BLAS/LAPACK interface, numeric autodiff, iterative solvers, numerical integration, arbitrary precision.

Build

Maintainers: the release procedure and the versioning rules are in RELEASING.md, beside this file.

make test        # build and run the test suite
make coverage    # line-coverage summary + HTML report
make sanitize    # tests under AddressSanitizer + UndefinedBehaviorSanitizer
make lint        # clang-tidy
make format      # uncrustify, in place
make doc         # API reference (Doxygen) with embedded coverage

Override the compiler with make test CXX=g++-14.

Supported toolchains

A C++20 compiler is required. These are the toolchains each push is built and tested with; an older one may work, and is not tested:

CMake 3.16 or newer; Python 3.10 or newer for the binding. The x86-64 kernels are checked on an emulated AVX-512 processor in CI.

numerical::numerical is the CMake target — add_subdirectory, FetchContent, or an installed config package:

# After `cmake --install <build> --prefix <prefix>`:
find_package(numerical CONFIG REQUIRED)
target_link_libraries(app PRIVATE numerical::numerical)

Python binding

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

A CPython binding (stable ABI, one cp310 abi3 wheel serves 3.10+) exposes the unifying object numerical.Tensor (rank-1 a vector, rank-2 a matrix, rank-0 a scalar; float64, complex128, float32 or complex64) and the algebra over it. Every function takes a Tensor or anything asarray reads — a NumPy array, a nested list, a number:

import numerical
a = [[1, 2], [3, 4]]
numerical.einsum("ij,jk->ik", a, [[5, 6], [7, 8]])   # matrix product
numerical.einsum("i,i->", [1, 2, 3], [1, 2, 3])       # dot
numerical.einsum("ij,jk", a, [[5, 6], [7, 8]])        # implicit output, as NumPy: "ij,jk->ik"
numerical.einsum("...ij,...jk->...ik", s, t)          # an ellipsis, its axes broadcast, as NumPy's
numerical.solve(a, [1.0, 1.0])                         # a @ x = b
numerical.fft.rfft([1.0, 2.0, 3.0, 4.0])               # numpy.fft's names: fft, ifft, fftn, rfft, irfft, fftfreq…
numerical.fft.dct([1.0, 2.0, 3.0, 4.0], norm="ortho")  # and scipy.fft's dct, idct, dst, idst, dctn…

A failure is a numerical.LinAlgError: a ValueError, as numpy.linalg.LinAlgError is, and a numerical.error. A Tensor indexes (t[i, j], t[1:5:2, ::-1], t[..., 0], t[None, i]), iterates, transposes (t.T) and multiplies (a @ b, numerical.matmul) as a NumPy array does, stacks of matrices along leading axes broadcast against each other, and computes as one:

t = numerical.arange(6).reshape(2, 3)
(t + 1) * [10, 0, 1] / 2      # element-wise, broadcast as NumPy's
t.sum(), t.mean(axis=0), t.max(axis=-1), t.sum(axis=(0, 1), keepdims=True)
numerical.zeros((2, 2)), numerical.eye(3), numerical.linspace(0, 1, 5)
numerical.exp(t), numerical.sqrt(t), abs(t), t ** 2

Shapes that do not broadcast raise a ValueError worded as NumPy's. max and min take real dtypes only (NumPy orders complex numbers lexicographically; NumericAL does not order them); max gives +0 and min −0 where both zeros are among the elements, as IEEE 754-2019 and NumPy on arm64 (NumPy on x86-64 keeps a zero's sign as its loops meet them). Operands of two dtypes promote as NumPy 2 promotes them: complex where one is, single precision only where both are, a Python number keeping the tensor's precision (float32 + 2.0 is float32). float32 and complex64 arithmetic is single precision, as NumPy's: + - * / and ** by an integer give its very bits, complex ones included (each fused as the NumPy beside it fuses it); a complex number's square root, exp, log, sin, cos, tan and tanh are glibc's, NumPy's on Linux to the bit (macOS's and Windows' libraries, NumPy's there, stray from them among infinities and NaNs and past their last bits), and its magnitude NumPy's own.

The decomposition layer is exposed too (a numpy.linalg-style surface): determinant(a) (det), slogdet(a) → (sign, logabsdet), inverse(a) (inv), pinv(a, rcond, hermitian), matrix_rank(a, tol, hermitian), eigvalsh(a, UPLO), lstsq(a, b, rcond) → (x, residuals, rank, s), cholesky(a, upper=False), qr(a, mode="reduced") → (Q, R) ("complete" as NumPy's, "r" → R alone), eigh(a, UPLO) → (values, vectors) (with b=, a positive-definite matrix, the pencil a x = λ b x as scipy.linalg.eigh(a, b); eigvalsh(a, b=b) its values), schur(a, output="real") → (T, Z) as scipy.linalg.schur ("complex" the triangular form of a real matrix), eig(a) → (values, vectors) and eigvals(a) of a general square matrix (LAPACK's dgeev/zgeev: balancing, blocked Hessenberg reduction, multishift QR with aggressive early deflation; complex answers whatever the eigenvalues, as NumPy 2.5's, where NumPy before 2.5 made them real where it could; the eigenvalues in the Schur form's order, a real matrix's pairs conjugate to the bit; each vector of unit norm, its largest component real), svd(a, full_matrices=True, compute_uv=True, hermitian=False) → (U, S, Vh) as NumPy's (full_matrices=False: the thin factors, so (U * S) @ Vh is a; compute_uv=False: S alone; hermitian=True: from the eigendecomposition, the singular values the eigenvalues' magnitudes), svdvals(a). The keywords mean what they mean to NumPy 2: UPLO="U" and upper=True read the upper triangle, rcond/rtol cut the singular values relative to the largest and tol absolutely, hermitian=True takes the eigendecomposition. Every one accepts the four dtypes and answers in the input's: a float32 or complex64 matrix is decomposed in double precision and its answers rounded to single, where NumPy's LAPACK computes in single (so they are as accurate as single precision holds, and the default tolerances are float32's epsilon, as NumPy's). eigh of a complex Hermitian matrix gives real values and complex vectors. From scipy.linalg, with its signatures: solve_triangular(a, b, trans=0, lower=False, unit_diagonal=False); lu_factor(a) → an LUFactor, the pair (lu, piv) SciPy returns (LAPACK's interchanges, a tuple of ints) with .solve(b), and lu_solve((lu, piv), b, trans=0); cho_factor(a, lower=False) → a CholeskyFactor, (c, lower), and cho_solve((c, lower), b): one factorisation, then n² k per solve. Their differences: a zero on U's diagonal raises in lu_solve, where SciPy's divides by it, and cho_factor leaves zeros in the triangle it does not use. Every one but lstsq (2-D in NumPy too) takes a stack of matrices, (..., m, n), as NumPy's: each matrix answers as it does alone, to the bit, and the answers stack along the leading axes (det and slogdet give tensors, a stack's matrix_rank nested integers); solve broadcasts the leading axes of a and b as NumPy 2 does, b a vector only when it is 1-D. Small matrices, or as many as threads, go to the threads one each; larger ones one at a time, each with the threads (10 000 inverses of order 4 on eight arm64 cores: 2.0 ms, NumPy 4.4 ms). eigh(a, method=...) and svd(a, method=...) choose the algorithm: divide and conquer by default (LAPACK's dsyevd/zheevd, dgesdd), "tridiagonal_ql" or "bidiagonal_qr" (dsyev's, dgesvd's), or "jacobi". The rest of numpy.linalg answers as NumPy 2's, on stacks too: norm(x, ord, axis, keepdims), vector_norm, matrix_norm ("fro", "nuc", 1, -1, 2, -2, ±inf), cond(x, p) (infinite for a singular matrix), matrix_power(a, n) (n negative through the inverse), multi_dot(arrays) (the cheapest order of the chain), trace(x, offset, dtype), diagonal(x, offset), matrix_transpose, outer, cross(x1, x2, axis), tensordot(a, b, axes), tensorsolve(a, b, axes), tensorinv(a, ind), vecdot(x1, x2, axis) (conjugating x1), and conj (conjugate). A 2-norm is scaled by its largest magnitude, so it neither overflows nor underflows: vector_norm([1e200, 1e200]) is 1.414e200, where NumPy's is infinite. A matrix holding a NaN or an infinity (in the triangle read, where only one is) raises numerical.LinAlgError, where NumPy returns NaN factors or raises depending on the routine: a decomposition of a non-finite matrix has no meaning, and saying so beats propagating NaN into every factor. A negative or NaN tolerance raises too, where NumPy's pinv and matrix_rank return infinities or a rank of 0.

A Tensor is a citizen of the scientific Python stack, zero-copy:

import numpy as np, torch
np.asarray(t)          # via __array_interface__ (a read-only view)
np.from_dlpack(t)      # via __dlpack__
torch.from_dlpack(t)   # the same DLPack capsule — JAX / CuPy consume it too
numerical.asarray(np.eye(3))   # the other direction (a copy)

asarray copies any NumPy array — bool, integer, float32/64 or complex64/128 elements, in either byte order, at any strides (a transposed, sliced or reversed view) — as well as a nested list or tuple of numbers and a plain number (rank 0): a float32, complex64 or complex128 array as it is, anything else into float64 (complex128 for a complex list or number), or into the dtype asked; a complex input is never cast to a real dtype. matmul and @ read a NumPy array where it lies, with no copy, when it is row-major, contiguous, aligned, in this machine's byte order and of one of the four dtypes, the other operand of the same dtype (a product of order 512 by a vector: 231 → 18 µs); solve, det, slogdet and inv read a square float64 or complex128 array so too, into the one copy their factorisation takes. Any other array is copied first, with the same answer. NumericAL makes no views: einsum("ij->ji", a) returns a new tensor where NumPy returns a view of a (a copy, faster than NumPy's copy of a.T), so writing to it never reaches a.

Compiled with GCC, the headers want -ffp-contract=on (Clang's default; the wheels and the sdist set it): GCC's own default, fast, fuses multiply-adds across statements as its passes find them, so a loop's vectorised and scalar versions, one or the other chosen at run time by the operands' addresses, can round an element differently, and the same operands then give other last bits from one call to the next.

Compiled with Clang for arm64, the headers want -mllvm -force-ordered-reductions=false too (the wheels and the sdist set it where the compiler takes it): Clang otherwise vectorises some loops that sum products in order, rounding their products apart where the same loop inlined elsewhere fuses them, so a sum's last bits depend on its caller (a row of a complex matrix times a vector against that row times the vector alone).

numerical.get_include() returns the header directory, so the C++ library can be located through its Python install. make python builds the extension in place; make python-test runs the binding suite — including a seeded differential fuzzer that checks hundreds of random einsum specifications and linear-algebra problems against NumPy within tolerance (when NumPy is importable, as in CI). PETSc (petsc4py) has no such exchange protocol — bridging it would require an explicit copy into a PETSc object.

BLAS and LAPACK for other programs

Experimental. The Linux wheels carry numerical_blas, a shared library of 39 BLAS and LAPACK routines under their Fortran names: ?gemm_, ?gesv_, ?getrf_, ?getrs_, ?potrf_ and ?potrs_ in the four precisions, sdot_, ddot_, cdotu_, cdotc_, zdotu_ and zdotc_, and dsyrk_, dgemv_, dtrsm_, dsyevd_, dgesdd_, dgeqrf_, dorgqr_, dgelsd_ and dgeev_ (gfortran's calling convention, a complex result returned by value); each with 32-bit integers, and under its name ending in 64_ (dgemm_64_) with 64-bit ones, as OpenBLAS's ILP64 build names them. numerical.blas_library() gives its path (None on macOS and Windows, where the wheels carry none: macOS's dyld binds each routine a program imports to the library that provides it, which a library inserted by DYLD_INSERT_LIBRARIES, DYLD_FORCE_FLAT_NAMESPACE or not, does not replace; there, CMake's build is linked by name, -lnumerical_blas). Preloaded, it answers those calls in a program linked against the system's BLAS and LAPACK, which keep every other routine:

LD_PRELOAD=$(python -c "import numerical; print(numerical.blas_library())") python script.py

Julia reaches the library through libblastrampoline, which loads it by its path and forwards it the routines it has (those it lacks stay OpenBLAS's). Twelve groups of Julia's LinearAlgebra tests pass with it on Linux, but for two that compare svd's vectors with gesvd's column by column: they agree up to the columns' signs, which LAPACK leaves open.

using LinearAlgebra
BLAS.lbt_forward(path; clear=false, suffix_hint="64_")  # path: numerical.blas_library()

The program must call BLAS by those names: a NumPy or SciPy from the system's packages (Debian, Ubuntu, Fedora…) or from FlexiBLAS, R, Octave, a Fortran or C program. NumPy and SciPy wheels from PyPI carry an OpenBLAS whose symbols are renamed, which a preloaded library does not reach. Importing numerical never loads it.

Its routines run NumPy's and SciPy's linear algebra test suites in CI, and Julia's (julia.yml, weekly and on a change to them); an argument is checked as the reference BLAS checks it (the reference XERBLA's message, then the call returns). A NaN or an infinity, which NumericAL's own functions refuse, is taken as the reference LAPACK (3.12) takes it, where it takes it one way: getrf, gesv, geqrf, orgqr and the solvers carry it through, potrf reports the order its factorisation stops at, syevd gives NaN eigenvalues and INFO n. gesdd, geev and gelsd report the matrix as an illegal argument (INFO -4), as the reference's gesdd and geev report a NaN; on an infinity the reference loops without end or gives NaN, and on a NaN its gelsd fails inside.

Threads and kernels

The kernels share a pool of threads, started the first time a kernel's work pays for more than one: one per physical core the process may run on, where the system tells them (Linux, macOS; elsewhere one per hardware thread), since two threads on one core's hyperthreads share its multiply-add units. numerical.set_num_threads(n) (C++: numerical::set_max_threads), or NUMERICAL_NUM_THREADS read at import, caps the threads a kernel uses, the caller's included, as OMP_NUM_THREADS caps BLAS; at 1 no thread is started. A process forked while the pool exists runs its kernels alone. Each library that embeds NumericAL's headers has a pool and a cap of its own.

float64 and complex128 matrix products, and the decompositions through them, run on a kernel chosen at run time, which numerical.gemm_path() (C++: gemm_path_in_use) reports:

• "portable" — the compiler's own vectors, on any processor;
• "neon" — on arm64, its vectors with each multiply-add fused, whatever the compiler would contract (float32 products too, there and on the AMX path, whose coprocessor takes float64 alone);
• "avx2", "avx512" — on x86-64, where cpuid reports the instructions and the system saves their registers (the wheels are built for the baseline x86-64 and carry these tiles beside it);
• "amx" — on Apple M1 to M4, the matrix coprocessor Apple's Accelerate uses, which Apple does not document; a product of few rows or columns, which it would compute as whole tiles of 32 × 16, runs on the NEON tile. It is used only on a known chip family, outside a virtual machine, and once a product computed on it equals std::fma in order to the bit; NUMERICAL_AMX=0, read by the Python package at import, keeps it off (its self-test included). C++ reads no environment: limit_gemm_path below amx before the first product does the same, and compiling with NUMERICAL_NO_AMX removes it.

numerical.limit_gemm_path(name) (C++: limit_gemm_path), or NUMERICAL_GEMM at import, caps the kernel at name or a more portable one. Every element of a product is summed in ascending order of the depth on any number of threads, so a result is the same, to the bit, whatever the thread count; across kernels it differs in the last bits, since the NEON, AVX and AMX kernels fuse each multiply-add and the portable one rounds as the compiler does (under the wheels' -ffp-contract=on, Clang fuses c + a·b there and GCC does not). Past the portable kernel, a complex product runs as one real product of twice the order, [Re A | −Im A] times [[Re B, Im B], [Im B, −Re B]] (by blocks), on that kernel and its threads: each element then sums the products of A's real parts, then those of its imaginary parts, where the portable kernel sums whole complex products in order.

Benchmarks

make bench-linalg times NumericAL against the BLAS/LAPACK NumPy ships, each library alone in its process, every answer checked first; make fuzz-linalg checks every decomposition against NumPy on adversarial matrices, on every kernel, and make fuzz-edges the element-wise operations and reductions on signed zeros, infinities, NaNs, values near overflow and empty axes. On an M1 Pro against Accelerate, at order 1024, matrix products, solve and inverse are as fast or faster (0.94×), Cholesky, QR and eigh within 1.0–1.2×, svd and lstsq 1.3–1.4×; small problems (order 64–256) take 1.5–3× Accelerate's time. Against OpenBLAS on x86-64 (i9-10900X, i9-9900K) large problems (order 512 and 1024) take 1.1–2.5×, eigh 2.6–2.9×. The tables, the method and where the gaps are: BENCHMARKS.md.

Numerical discipline

Floating-point accuracy is documented, never claimed exact where it cannot be: every decomposition is backward stable, checked against its defining property and against NumPy within bounds of the order times the machine epsilon. A result is the same, to the bit, on any number of threads; between kernels and processors it may differ in the last bits (see Threads and kernels). Same gate bar as the rest of the ecosystem: 100% coverage, clang and g++-14, lint, format, MISRA-oriented analysis and sanitizers green.

Accuracy, function by function

The bounds the differential fuzzer (make fuzz-linalg, and every week fuzz-long.yml) holds every answer to, on adversarial matrices (clustered, repeated and graded spectra, rank deficiency, exact integers, entries scaled to 1e±150, complex data) at every kernel the processor has. Throughout, n is the larger dimension, ε = 2⁻⁵³, ‖·‖ the Frobenius norm and c = 50 n ε: a backward-stable algorithm's error is a small multiple of n ε whatever the conditioning, and a wrong update or deflation misses it by orders of magnitude.

License

MIT — see LICENSE. Several algorithms follow LAPACK's routines closely; its BSD license, which their redistribution carries, is in NOTICE.

Author

René Chenard