Struct M16TcDiff 

pub struct M16TcDiff {
    pub max_ulp: i32,
    pub over_budget: Vec<M16TcOutlier>,
    pub over_ulp_only: usize,
    pub sign_flips: usize,
    pub max_abs: f64,
    pub rel_rms: f64,
    pub rms: f64,
    pub row_rms: Vec<f64>,
}

The result of comparing an m x n BF16 block against the scalar reference.

Fields

max_ulp: i32

Largest ordinal BF16 ULP distance, sign flips excluded and counted.

over_budget: Vec<M16TcOutlier>

Elements the FULL criterion rejected — ordinal budget AND the accumulation floor. This is the number that gates a cell.

over_ulp_only: usize

Elements the ordinal budget alone would have rejected. Reported so a round-6-style cell can be read at a glance as “cancellation tail” rather than investigated as a defect.

sign_flips: usize

max_abs: f64

rel_rms: f64

rms: f64

RMS of the whole REFERENCE block. The floor is per-ROW, but this is the number the receipts print and the one that makes an outlier’s magnitude interpretable at a glance.

row_rms: Vec<f64>

Per-row reference RMS, in row order — the scales the floor was actually evaluated at. Printed by the oracles next to each outlier.

Trait Implementations

impl Clone for M16TcDiff

fn clone(&self) -> M16TcDiff

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for M16TcDiff

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

Formats the value using the given formatter.

impl Default for M16TcDiff

fn default() -> M16TcDiff

Returns the “default value” for a type.