Struct MinFill 

pub struct MinFill;

Approximate Minimum Fill metric (HAMF4) — Amestoy 1999 thesis.

AMF selects the next pivot to minimise the fill introduced by the elimination, rather than the candidate’s degree. On bipartite- KKT graphs with a few “hub” rows AMF can be 47× better than AMD on final nnz_L (see dev/research/amf-clean-room.md Section 1).

Score is a quantized RMF = DEG*(DEG-1+2*DEGME) - WF(i) value stored in i32. Buckets up to and including NORIG = n are one bucket per integer score; above NORIG the buckets quantize with stride PAS = max(n / 8, 1) and the head must be linear-scanned to pick the minimum-RMF entry. Supervariable absorption merges the per-supervariable WF with max.

Inner loop: MinFill::run_elimination dispatches to run_elimination_amf (Phase B.2 of dev/plans/amf-clean-room.md). The lazy WF(e) cache, three-accumulator Pass-2, supervariable max-merge of wf, saturated/regular RMF branch, and coarse-bucket linear scan all live in algo.rs.

Trait Implementations

impl Clone for MinFill

fn clone(&self) -> MinFill

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for MinFill

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

Formats the value using the given formatter.

impl Default for MinFill

fn default() -> MinFill

Returns the “default value” for a type.

impl Metric for MinFill

fn n_buckets(n: usize) -> usize

AMF needs 2 * n + 2 slots: 0..=NORIG for one-per-score fine buckets, NORIG+1..=NBBUCK (NBBUCK = 2 * n) for coarse-stride buckets, and one halo slot at NBBUCK + 1 reserved for V1 boundary variables (inert in our use case; see Section 11 of dev/research/amf-clean-room.md).

fn init_score(len: i32) -> i32

Seed the per-supervariable score with the row’s adjacency length len(i). Same seeding as AMD; the AMF metric only diverges once the elimination loop starts producing elements.

fn bucket(score: i32, n: usize) -> usize

Quantize score into a bucket index.

0..=n are fine buckets (one per integer score). Above n the buckets are coarse with stride PAS = max(n / 8, 1), capped at NBBUCK = 2 * n. Negative scores (which should not occur in a well-formed AMF run, but we defend against truncation underflow on RMF / (NVI + 1)) clamp to bucket 0.

Reference: ana_orderings.F:4954-5017 and dev/research/amf-clean-room.md Section 4.

fn coarse_bucket(idx: usize, n: usize) -> bool

Coarse buckets are those above NORIG = n. select_pivot must walk the bucket chain and pick the entry with the smallest exact score (ana_orderings.F:4392-4418).

fn merge_supervariable(parent: &mut i32, child: i32)

On supervariable merge j → i, update the surviving anchor’s score with max(WF(i), WF(j)) (ana_orderings.F:4920).

type Score = i32

Bucket key produced by the selection metric. AMD uses i32 (the running degree); AMF will also use i32 (quantized RMF).

fn run_elimination( ws: &mut Workspace, aggressive: bool, ) -> Result<StepFlops, OrderingError>

Run the metric’s elimination loop on a freshly initialised Workspace. Returns the accumulated flop counters.

impl Copy for MinFill