Struct TensorCoreEngine 

pub struct TensorCoreEngine { /* private fields */ }

Tensor Core MMA engine.

Provides mma() (D = A·B + C) matching the semantics of CUDA’s nvcuda::wmma::mma_sync and PTX mma.sync instructions.

Implementations

impl TensorCoreEngine

pub fn new(precision: MmaPrecision, shape: FragmentShape) -> Self

Create a new engine with specified precision and fragment shape.

pub fn mma(&self, a: &Fragment, b: &Fragment, c: &Fragment) -> Result<Fragment>

Matrix multiply-accumulate: D = A · B + C

A: (m × k), B: (k × n), C: (m × n) → D: (m × n)

pub fn gemm( &self, a: &[f32], b: &[f32], c: &mut [f32], m: usize, n: usize, k: usize, alpha: f32, beta: f32, ) -> Result<GemmStats>

Full GEMM using tiled MMA: C = alpha * A · B + beta * C

A: (m × k), B: (k × n), C: (m × n) — arbitrary sizes, tiled internally.