Enum ArrowSolverMode 

pub enum ArrowSolverMode {
    Direct,
    SqrtBA,
    InexactPCG,
}

BA Schur solve variant for the reduced shared β system.

• ArrowSolverMode::Direct is BA’s dense reduced-camera-system solve: eliminate the per-point/per-row blocks, form the reduced system, and Cholesky factor it. This is the Ceres/g2o default for modest camera counts and is appropriate here for K <= 2000. GPU support: ✓ — requires dense H_ββ and dense per-row H_tβ slabs.

• ArrowSolverMode::SqrtBA ports Square-Root BA (Demmel/Gao/Gu et al., CVPR 2021): Schur terms are formed as (L_i^-1 H_tβ_i)^T (L_i^-1 H_tβ_i) from the per-row square-root factor L_i, avoiding explicit H_tt^-1 H_tβ products. It is the preferred direct path when single-precision assembly is introduced or when row blocks are poorly conditioned. GPU support: ✓ — requires dense H_ββ and dense per-row H_tβ slabs.

• ArrowSolverMode::InexactPCG ports “Bundle Adjustment in the Large” (Agarwal et al.): the Schur system is solved inexactly by PCG with a Jacobi Schur preconditioner, avoiding dense K × K factorization for SAE-manifold scale shared systems. GPU support: CPU only until the row-procedural H_tβ GPU PCG path (issue #288 Part B) is wired. The topology selector must not request InexactPCG via the GPU entry point; solve_arrow_newton_step returns GpuRequiresDenseSystem for matrix-free systems, and the wrapper in solver/gpu/arrow_schur_gpu.rs routes those to CPU InexactPCG automatically. At K ≥ 5000 the GPU PCG path will supersede the CPU path once the row-procedural H_tβ kernel and boxed GPU matvec backend in run_pcg_with_preconditioner are wired.

Variants

Direct

SqrtBA

InexactPCG

Implementations

impl ArrowSolverMode

pub const fn automatic(k: usize) -> Self

BA-size heuristic: dense RCS for modest K, inexact Schur PCG for large shared systems. This follows Agarwal et al.’s direct-vs-iterative split for large BA, mapped from cameras to decoder coefficients.

pub const fn automatic_for_single_precision(k: usize) -> Self

Square-Root BA is the direct-solve stability mode for future f32 callers. Large K still routes to inexact PCG because dense Schur storage dominates precision concerns at that scale.

Trait Implementations

impl Clone for ArrowSolverMode

fn clone(&self) -> ArrowSolverMode

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for ArrowSolverMode

impl Debug for ArrowSolverMode

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

Formats the value using the given formatter.

impl Eq for ArrowSolverMode

impl PartialEq for ArrowSolverMode

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

Tests for self and other values to be equal, and is used by ==.

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

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for ArrowSolverMode