src/uni20/linalg/backends/cpu

This directory contains CPU dense linalg implementations that do not require a vendor accelerator backend.

Contents

• contract.hpp: host lease acquisition and dispatch for arbitrary-rank pairwise contraction.
• copy.hpp: reference accessor-respecting element-copy kernel.
• conjugate_inplace.hpp: reference accessor-respecting in-place conjugation kernel.
• transform.hpp: generic accessor-respecting elementwise overwrite and update kernels.
• strided_transform.hpp: internal CPU executor for compact strided iteration plans shared by copy, conjugation, and generic transforms.
• gemm.hpp: reference accessor-respecting GEMM kernel.
• gemv.hpp: reference accessor-respecting GEMV kernel.
• reductions.hpp: shared accessor-respecting reduction traversal, full and partial sums, inner products, and scaled sum-of-squares Euclidean norms.
• matrix_norm.hpp: accessor-respecting maximum-entry, induced one/infinity, and Frobenius matrix norms.
• linear_solve.hpp: accessor-respecting Gaussian elimination with partial pivoting over destructive matrix workspaces.
• matrix_set.hpp: accessor-respecting structured matrix initialization.
• matrix_exponential.hpp, matrix_exponential.cpp: adaptive dense matrix exponential over uni20::DenseMatrix, composed from dispatched matrix products, transforms, norms, structured initialization, and solves.
• `detail/`: shared same-precision numerical accumulation helpers.

Notes

• This is the home for CPU-specific dense algorithms that are above raw kernels but below public tensor workflows.
• Reference elementwise kernels prioritize direct, deterministic semantics. Layout-specific tiling, vectorization, and HPTT-like traversal belong in separate optimized backends that may decline unsupported instances.
• Keep matrix-free algorithms in the Krylov source layer; only small dense projected problems or direct dense linalg belong here.

Related Documentation

• Linalg backend source map
• Linear algebra documentation
• Krylov algorithms