Overview

template<typename T>
struct batchlas::GesvdjParams< T >

Parameters for the one-sided Jacobi SVD (gesvdj_cta).

Trap: Deliberately NOT a reuse of JacobiParams: its sort_order defaults to Ascending while gesvd admits exactly one order (descending), so reuse invites a silent flip.

Definition at line 1731 of file extensions.hh.

#include <batchlas/blas/extensions.hh>

Types

◆ Real

template<typename T>

using batchlas::GesvdjParams< T >::Real = typename base_type<T>::type

Definition at line 1732 of file extensions.hh.

Variables

◆ tol_multiplier

template<typename T>

Real batchlas::GesvdjParams< T >::tol_multiplier = Real(1)

A rotation is applied to pivot pair (p,q) only when \( |a_{pq}| > \mathrm{tol\_multiplier} \cdot n \varepsilon \sqrt{|a_{pp}| |a_{qq}|} \) over the 2x2 Gram entries of the current columns; again the relative test.

Definition at line 1737 of file extensions.hh.

◆ max_sweeps

template<typename T>

size_t batchlas::GesvdjParams< T >::max_sweeps = 30

Cap on cyclic sweeps; convergence normally takes well under 10.

Definition at line 1739 of file extensions.hh.

◆ cta_wg_size_multiplier

template<typename T>

size_t batchlas::GesvdjParams< T >::cta_wg_size_multiplier = 1

Multiplies the baseline problems-per-work-group.

Baseline is 32 / P problems, clamped by local memory and max work-group size.

Definition at line 1743 of file extensions.hh.

◆ zero_sigma_multiplier

template<typename T>

Real batchlas::GesvdjParams< T >::zero_sigma_multiplier = Real(1)

\( \sigma_j \le \mathrm{zero\_sigma\_multiplier} \cdot \varepsilon \sigma_{\max} \) means \( U_j \) is not determined by A and is filled from the orthogonal complement.

Definition at line 1747 of file extensions.hh.

◆ sweep_counts

template<typename T>

Span<int32_t> batchlas::GesvdjParams< T >::sweep_counts = Span<int32_t>()

Optional per-problem diagnostic, the analogue of cusolverDnXgesvdjGetSweeps: when non-empty (size >= batch_size) the kernel writes each problem's sweeps.

Definition at line 1751 of file extensions.hh.