Trait Gemm 

pub trait Gemm:
    Copy
    + Default
    + Sealed
    + 'static {
    // Required methods
    fn mul_add(acc: Self, a: Self, b: Self) -> Self;
    unsafe fn gemm(
        m: usize,
        n: usize,
        k: usize,
        dst: &mut [Self],
        a: MatView<'_, Self>,
        b: MatView<'_, Self>,
    );
}

An element type with a GEMM kernel: f32, f64, f16, and [f32; 2] (complex64 as (re, im)).

Required Methods

fn mul_add(acc: Self, a: Self, b: Self) -> Self

acc + a·b, used by the reference loop (and under Miri).

unsafe fn gemm( m: usize, n: usize, k: usize, dst: &mut [Self], a: MatView<'_, Self>, b: MatView<'_, Self>, )

Calls gemm::gemm for dst (row-major m×n) = lhs · rhs.

Safety

Every element reachable through the views must lie inside their data, and dst must hold m·n elements.

Dyn Compatibility

This trait is not dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementations on Foreign Types

impl Gemm for [f32; 2]

fn mul_add(acc: Self, a: Self, b: Self) -> Self

unsafe fn gemm( m: usize, n: usize, k: usize, dst: &mut [Self], a: MatView<'_, Self>, b: MatView<'_, Self>, )

impl Gemm for f16

unsafe fn gemm( m: usize, n: usize, k: usize, dst: &mut [Self], a: MatView<'_, Self>, b: MatView<'_, Self>, )

As MLX: widened to f32, multiplied with f32 accumulation, rounded once (the gemm crate’s f16 kernel accumulates in f16, losing bits over a long k).

fn mul_add(acc: Self, a: Self, b: Self) -> Self

impl Gemm for f32

fn mul_add(acc: Self, a: Self, b: Self) -> Self

unsafe fn gemm( m: usize, n: usize, k: usize, dst: &mut [Self], a: MatView<'_, Self>, b: MatView<'_, Self>, )

impl Gemm for f64

fn mul_add(acc: Self, a: Self, b: Self) -> Self

unsafe fn gemm( m: usize, n: usize, k: usize, dst: &mut [Self], a: MatView<'_, Self>, b: MatView<'_, Self>, )

Implementors