Struct IntGemmArgs 

pub struct IntGemmArgs<'a, T, BT = f32>
where
    T: IntElement,
    BT: BiasElement,
{
    pub a: MatrixRef<'a, T>,
    pub b: MatrixRef<'a, T>,
    pub c: Option<MatrixRef<'a, T>>,
    pub d: MatrixMut<'a, T>,
    pub bias: Option<VectorRef<'a, BT>>,
    pub alpha: f32,
    pub beta: f32,
}

Per-launch arguments for an IntGemmPlan::run call.

Parallel to GemmArgs for the integer GEMM family. The matrix operands carry the kernel element type T: IntElement (today: S8 or U8); the optional bias carries the independent bias element type BT: BiasElement (today: f32 or i32). Scalar alpha / beta are always f32 regardless of T or BT — CUTLASS’s LinearCombinationClamp / LinearCombinationBiasElementwise epilogues do the entire alpha/beta/bias/activation chain in float (after int32→float dequant of the accumulator) and saturating-cast back to the int output range on store.

Fields

a: MatrixRef<'a, T>

Left input. Row-major [M, K].

b: MatrixRef<'a, T>

Right input. Column-major [K, N] (RCR).

c: Option<MatrixRef<'a, T>>

Optional accumulation source. Row-major [M, N].

d: MatrixMut<'a, T>

Output. Row-major [M, N].

bias: Option<VectorRef<'a, BT>>

Optional bias vector. Required when the descriptor’s epilogue is any Bias* variant; must be None for EpilogueKind::Identity. Length-N, contiguous (stride 1) device memory; broadcast across rows of D.

alpha: f32

Multiplier on the matrix-multiply accumulator. Always f32 for int GEMM — CUTLASS does the entire epilogue compute in float space.

beta: f32

Multiplier on c. Forced to 0 internally when c is None.

Trait Implementations

impl<'a, T, BT> Debug for IntGemmArgs<'a, T, BT>

where T: Debug + IntElement, BT: Debug + BiasElement,

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

Formats the value using the given formatter.