pub fn compute_sve_dense<K>(
kernel: K,
epsilon: Option<f64>,
twork: TworkType,
full_domain: bool,
) -> Result<SVEResult, Error>where
K: ExpandableKernel,Expand description
compute_sve by the dense CPQR-based SVD of the fully formed blocks
(SVEStrategy::matrices and compute_svd_dtensor), as before the ACA
(and as compute_sve still does in f64): the reference the ACA path is
tested against
It forms every block and decomposes it in full, which in Float64X2 takes
far longer than compute_sve (about 50 s, on 8 threads, for the kernel
e^{-Λy(x+1)/2} / (1 + e^{-Λy}) with Λ = 10⁶), so it is meant for checking
the result of compute_sve on a new kernel. The cutoff is the default one of
compute_sve, 2 * machine epsilon.
With full_domain, the strategy of SVEOptions::full_domain
(NonCentrosymmSVE) instead of that of the kernel.
§Errors
The errors of compute_sve