Struct Matrix 

pub struct Matrix { /* private fields */ }

Dense matrix stored in row-major format

The matrix data is stored as a flat vector where each row is laid out contiguously in memory for cache efficiency. This layout enables efficient SIMD operations on individual rows.

Fields

• n_col - Number of elements per row (dimension of embeddings)
• data - Flat storage of matrix data (rows * ncol elements)

Implementations

impl Matrix

pub fn new(n_row: usize, n_col: usize) -> Matrix

Creates a new matrix initialised with zeros

Params

• rows - Number of rows in the matrix (typically vocabulary size)
• n_col - Number of elements per row (embedding dimension)

Returns

A new Matrix with all elements set to 0.0

pub fn norm_self(&mut self)

Normalises all rows in the matrix to unit length

Each row vector is divided by its L2 norm, making it a unit vector.

pub fn make_send(self) -> MatrixWrapper

Wraps the matrix in a thread-safe wrapper

Returns

A MatrixWrapper that can be safely shared across threads using Arc

pub fn uniform(&mut self, bound: f32, seed: usize)

Initialises matrix with uniform random values

Params

• bound - Values will be sampled uniformly from [-bound, bound]
• seed - Seed for reproducibility.

pub fn norm(&self, i: usize) -> f32

Computes the L2 norm of a matrix row

Params

• i - Row index

Returns

The L2 norm (Euclidean length) of the row vector

pub fn zero(&mut self)

Sets all matrix elements to zero

pub unsafe fn add_row(&mut self, vec: *const f32, i: usize, mul: f32)

Adds a scaled vector to a matrix row (SAXPY operation)

Performs: row[i] = row[i] + mul * vec

This is used during gradient updates where we add scaled gradients to embedding vectors.

Params

• vec - Pointer to the vector to add (must have n_col elements)
• i - Row index
• mul - Scaling factor for the vector

Safety

The caller must ensure vec points to at least n_col valid f32 elements.

pub unsafe fn dot_row(&self, vec: *const f32, i: usize) -> f32

Computes dot product between a vector and a matrix row

Params

• vec - Pointer to the vector (must have row_size elements)
• i - Row index

Returns

The dot product result

Safety

The caller must ensure vec points to at least row_size valid f32 elements

pub fn dot_two_row(&self, i: usize, j: usize) -> f32

Computes dot product between two matrix rows

Params

• i - First row index
• j - Second row index

Returns

The dot product of row i and row j

pub fn get_row(&mut self, i: usize) -> *mut f32

Gets a mutable pointer to a matrix row

Params

• i - Row index

Returns

Mutable pointer to the start of row i

Safety

The caller must ensure the pointer is used correctly and doesn’t create aliasing issues. Primarily used for passing to SIMD functions.

pub fn row_as_slice(&self, i: usize) -> &[f32]

Returns a shared slice of row i

Params

• i - Row index

Returns

Slice of the row data

pub fn get_row_unmod(&self, i: usize) -> *const f32

Gets a const pointer to a matrix row

Params

• i - Row index

Returns

Const pointer to the start of row i

pub fn n_col(&self) -> usize

Returns the number of elements per row

pub fn n_rows(&self) -> usize

Returns the total number of rows

pub fn to_faer(&self) -> Mat<f32>

Converts the matrix to a Faer matrix

Returns

Faer matrix

pub fn write_csv(&self, path: &str) -> Result<()>

Write the matrix rows to a CSV file

Params

• path - Path to the output CSV

pub fn average_with(&self, other: &Matrix) -> Matrix

Compute the element-wise average of two matrices

Params

• other - The other matrix (must have same dimensions)

Returns

A new matrix where each element is (self + other) / 2

Panics

Panics if dimensions do not match.

Trait Implementations

impl Debug for Matrix

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

Formats the value using the given formatter.