Module cpu_kernels 

Injectable CPU linear-algebra kernels.

A CpuLinalgKernels implementation replaces tenferro-linalg’s built-in faer/LAPACK kernels op by op on one CPU backend. Install it with install_linalg_kernels on a CpuProviderBundleBuilder; the CPU session then asks it first for every primitive it dispatches (Cholesky, triangular solve, LU and full-pivot LU, solve, SVD, QR and rank-revealing QR, eigh, eig, and their value-only forms, including the _read variants). A method returns CpuLinalgOutcome::Unsupported, its default, before producing anything, and the built-in kernel runs instead. The Householder family, lu_factor and prepared LU solves, and _into outputs always use the built-in kernels. Composites (det, slogdet, inv, lstsq, pinv, norms) are built from the primitives and follow them.

A kernel runs inside the session’s entered CpuExecutionContext on host operands and must return exactly what the built-in kernel returns for the same input: the same outputs in the same order, shapes, dtypes, pivot and ordering conventions, and batch handling (trailing batch axes). It may run in parallel only in tenferro_cpu::ParallelMode::Inner, within thread_budget(), on the context’s pool (CpuExecutionContext::rayon_pool).

Examples

use std::sync::Arc;
use tenferro_cpu::{CpuBackendKind, CpuProviderBundle};
use tenferro_linalg::cpu_kernels::{install_linalg_kernels, CpuLinalgKernels, CpuLinalgKernelsSlot};

/// Declines everything, so the built-in kernels run.
#[derive(Debug)]
struct Nothing;
impl CpuLinalgKernels for Nothing {}

let builder = CpuProviderBundle::builder(CpuBackendKind::default_compiled());
let bundle = install_linalg_kernels(builder, Arc::new(Nothing)).build()?;
// Install with `CpuBackend::with_provider_bundle` (a faer backend; explicit
// bundles on the BLAS kind are rejected because BLAS threading is not
// enforceable).
assert!(bundle.extension::<CpuLinalgKernelsSlot>().is_some());

Structs

CpuLinalgKernelsSlot

The provider-bundle extension that carries a CpuLinalgKernels.

TriangularSolveOptions

Flags of a triangular solve (op(A) X = B or X op(A) = B).

Enums

CpuLinalgOutcome

Result of a CpuLinalgKernels call.

Traits

CpuLinalgKernels

Replacement CPU linear-algebra kernels. Every method defaults to CpuLinalgOutcome::Unsupported; implement the ones the provider handles. See the module documentation for the contract.

Functions

install_linalg_kernels

Install kernels on a provider bundle under construction.