Enum SvdDriver 

pub enum SvdDriver {
    Auto,
    Gesvdj,
    Gesvd,
    Xgesvdp,
}

cuSOLVER SVD routine selection used by SvdOptions on the CUDA backend.

The driver changes speed and numerical behavior; Self::Xgesvdp may perturb near-singular inputs. Gauge handling and AD rule selection are unchanged. CPU providers have a single SVD kernel and ignore the driver. Equivalent to jax.lax.linalg.svd(..., algorithm=...) and torch.linalg.svd(..., driver=...).

Examples

use tenferro_linalg::{SvdDriver, SvdOptions};

let options = SvdOptions::default().driver(SvdDriver::Gesvd);
assert_eq!(options.driver, SvdDriver::Gesvd);
assert_eq!(SvdOptions::default().driver, SvdDriver::Auto);

Variants

Auto

The backend’s default policy. On CUDA this is the accuracy-preserving, JAX-compatible rule: gesvdj when both matrix dimensions are at most 1024, otherwise gesvd. It never selects the polar Self::Xgesvdp, whose small singular values are only accurate to about 1e-6 of the largest (#2182). CPU providers have a single SVD kernel and ignore the driver.

Gesvdj

cuSOLVER’s Jacobi driver (cusolverDn<t>gesvdj) regardless of size.

Gesvd

cuSOLVER’s QR-based driver (cusolverDn<t>gesvd) regardless of size.

Xgesvdp

cuSOLVER’s polar-decomposition driver (cusolverDnXgesvdp).

Explicit only: Self::Auto never selects this routine. It is the fastest of the three drivers for min(m, n) >= 256 on an A100 (see the cuda_svd_driver benchmark), but its small singular values are only accurate to about 1e-6 of the largest singular value. Do not use it for an eps-scale truncation (a pseudoinverse or tensor-network truncation at 1e-12..1e-14): the perturbation can change the retained rank.

Tenferro reads cuSOLVER’s h_err_sigma perturbation bound and returns a typed backend failure instead of silently wrong factors when it is not small relative to the largest singular value. On rank-deficient F32/C32 input cuSOLVER reports a perturbation comparable to the spectrum, so an explicit single-precision call there fails rather than returning the perturbed values. CPU providers accept and ignore this selection.

Accuracy is input-dependent, not uniformly better or worse than Self::Gesvd. On ten-decade spectra built as U diag(s) Vᴴ it was the most accurate of the three drivers on an A100, but on a circulant-like matrix with the same spectrum it lost about two digits against gesvd. Measure on your own inputs before relying on it for small singular values.

Trait Implementations

impl Clone for SvdDriver

fn clone(&self) -> SvdDriver

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for SvdDriver

impl Debug for SvdDriver

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for SvdDriver

fn default() -> SvdDriver

Returns the “default value” for a type.

impl Eq for SvdDriver

impl PartialEq for SvdDriver

fn eq(&self, other: &SvdDriver) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for SvdDriver