Enum MmaShape 

pub enum MmaShape {
    M16N8K16,
    M16N8K32,
}

The shape of an mma.sync instruction.

Each variant corresponds to a distinct hardware tile geometry. Adding a new shape requires a new enum variant and a new set of fragment types because the per-thread register distribution is shape-dependent. See crate::fragment.

Variants

M16N8K16

16×16 × 16×8 → 16×8, fp16 / bf16 inputs with fp32 accumulate (Ampere+).

M16N8K32

16×32 × 32×8 → 16×8, signed int8 inputs with int32 accumulate (Ampere+).

Used by mma.sync.aligned.m16n8k32.row.col.s32.s8.s8.s32. K is 32, not 16 — twice the K-tile of the fp16 path. Introduced in Sprint 7.1 for INT8 dequantize-matmul.

Implementations

impl MmaShape

pub fn ptx_token(&self) -> &'static str

PTX shape token (e.g. "m16n8k16").

pub fn min_sm(&self) -> u32

Minimum SM version required to execute this shape.

Used by crate::ir::PtxModule::validate to reject kernels that emit tensor-core ops against too-low a target SM.

Trait Implementations

impl Clone for MmaShape

fn clone(&self) -> MmaShape

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for MmaShape

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

Formats the value using the given formatter.

impl PartialEq for MmaShape

fn eq(&self, other: &MmaShape) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Copy for MmaShape

impl Eq for MmaShape

impl StructuralPartialEq for MmaShape