Struct TritFloatTensor 

pub struct TritFloatTensor {
    pub data: Vec<TritFloat>,
    pub shape: Vec<usize>,
}

An N-dimensional tensor of TritFloats laid out in row-major order.

Each element carries its own confidence field. The tensor’s overall confidence is determined by min_confidence() or mean_confidence().

Fields

data: Vec<TritFloat>

shape: Vec<usize>

Implementations

impl TritFloatTensor

pub fn zeros(shape: &[usize]) -> Self

All-zero tensor with neutral confidence (0.5) at every element.

pub fn ones(shape: &[usize]) -> Self

All-ones tensor with maximum confidence at every element.

pub fn from_f32_slice(data: &[f32], shape: &[usize]) -> Self

Build from a flat f32 slice, all elements get confidence=1.0.

pub fn from_f32_with_confidence( vals: &[f32], conf: &[f32], shape: &[usize], ) -> Self

Build from f32 values with per-element confidence.

pub fn from_tritmatrix(m: &TritMatrix) -> Self

Convert a TritMatrix to a 2D TritFloatTensor.

Weights are exactly {-1, 0, +1} — they carry maximum confidence (1.0).

pub fn shape(&self) -> &[usize]

pub fn ndim(&self) -> usize

pub fn numel(&self) -> usize

pub fn get(&self, idx: &[usize]) -> TritFloat

pub fn set(&mut self, idx: &[usize], val: TritFloat)

pub fn matmul(a: &Self, b: &Self) -> Self

2D matrix multiply: [m, k] × [k, n] → [m, n].

Each output element’s confidence = min confidence over all contributing (a_i × b_i) multiplications. Zero-phase activations skip their MAC (@sparseskip at activation level) but still participate in the confidence running-minimum so uncertain zeros don’t disappear. Rows are computed in parallel via Rayon.

pub fn matmul_sparse(a: &Self, b: &Self) -> (Self, usize)

Matmul returning (result, total_macs_skipped) for sparsity instrumentation.

pub fn matmul_trit(activations: &Self, weights: &TritMatrix) -> (Self, usize)

Multiply float activations by exact ternary weights (TritMatrix).

This is the inference-time hot path: activations carry live confidence, weights are exact {-1,0,+1} with confidence=1.0, so output confidence = min activation confidence over each dot product.

@sparseskip fires on BOTH activation zeros (is_zero()) AND weight zeros (w == 0), giving the full combined sparsity savings.

Returns (output_tensor, total_macs_skipped).

pub fn add_elementwise(a: &Self, b: &Self) -> Self

pub fn mul_elementwise(a: &Self, b: &Self) -> Self

pub fn map<F>(&self, f: F) -> Self

where F: Fn(TritFloat) -> TritFloat + Sync + Send,

Apply a function to every element in parallel.

pub fn sum_all(&self) -> TritFloat

pub fn mean_all(&self) -> TritFloat

pub fn min_confidence(&self) -> f32

Minimum confidence across all elements: how certain is the least-certain value?

pub fn mean_confidence(&self) -> f32

Mean confidence: average epistemic certainty across the tensor.

pub fn confidence_histogram(&self) -> [usize; 9]

Histogram of confidence states across all elements. Index i = count of elements with confidence ≈ i/8. The 9 bins correspond to the 9 discrete states of the 2-trit confidence field.

pub fn sparsity(&self) -> f64

Fraction of elements with zero phase (exactly 0.0 value).

pub fn to_f32_vec(&self) -> Vec<f32>

Extract f32 values, discarding confidence information.

pub fn to_tritmatrix(&self) -> TritMatrix

Quantize to TritMatrix: positive phase → Affirm, negative → Reject, zero → Tend.

pub fn softmax_rows(&self) -> Self

Apply softmax along each row of a 2D tensor.

Trait Implementations

impl Clone for TritFloatTensor

fn clone(&self) -> TritFloatTensor

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TritFloatTensor

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

Formats the value using the given formatter.