Enum MatmulPrecision 

pub enum MatmulPrecision {
    Auto,
    Int8Dp4a,
    Int64,
    Bf16Hmma,
    Fp16Hmma,
}

Contraction precision for a lowered matmul.

Variants

Auto

Best general case: pick the always-correct path unless the value analysis proves a faster one is safe. Today resolves to ResolvedPrecision::Int64 (we do not yet carry a proof that operands fit int8 through the recognizer), so Auto is exactly “the correct-for-any-program default”.

Int8Dp4a

Signed-int8 __dp4a accumulation — faster, but VALID ONLY when operands provably fit int8 and the contraction fits int32. Not strictly better (restricted domain) → a knob, off by default.

Int64

64-bit integer accumulation — always correct (matches the VM’s i64 semantics for any operands), slower.

Bf16Hmma

bf16 tensor-core GEMM (HMMA 16×16×16, bf16 operands, f32 accumulate) — the half-precision fast path for a matmul over Real tensors on a 16-aligned shape (sm_80+). When the block tile divides the shape and its footprint fits static shared memory, the lowering auto-selects the cp.async double-buffered kernel (the measured 1.43× lever, select_hmma_kernel); otherwise the basic one-warp WMMA kernel. On an incapable shape/arch it falls back to the always-correct int64 kernel. A knob (restricted domain), off by default.

Fp16Hmma

fp16 tensor-core GEMM (HMMA 16×16×16, f16 operands, f32 accumulate) — the fp16 counterpart (sm_70+).

Implementations

impl MatmulPrecision

pub fn resolved(self) -> ResolvedPrecision

Resolve the knob to a concrete lowering. Auto conservatively picks the always-correct Int64 — we only emit a fast path (Int8Dp4a/Bf16Hmma/Fp16Hmma) when the caller explicitly requests it, which is precisely “default to best general; opt into the trade-off”.

Trait Implementations

impl Clone for MatmulPrecision

fn clone(&self) -> MatmulPrecision

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for MatmulPrecision

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

Formats the value using the given formatter.

impl PartialEq for MatmulPrecision

fn eq(&self, other: &MatmulPrecision) -> 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 MatmulPrecision

impl Eq for MatmulPrecision

impl StructuralPartialEq for MatmulPrecision