Struct ComputeKnobs 

pub struct ComputeKnobs {
    pub fidelity: MathFidelity,
    pub fp32_dest_acc: bool,
    pub packer_l1_acc: bool,
    pub pcc_floor: f64,
}

The compute NUMERICS knobs of an emitted matmul program. They are part of the program’s identity: the host’s ComputeConfig carries them, the program prints them on its LOGOS_MM_PASS line, and the ledger refuses a row whose printed knobs differ from what the harness requested — so a LoFi run can never again be compared to a HiFi4 run as if they were one config.

Fields

fidelity: MathFidelity

FPU passes per multiply (see MathFidelity).

fp32_dest_acc: bool

Accumulate in fp32 in the DST register file (halves the DST tile budget; higher PCC).

packer_l1_acc: bool

Accumulate K-block partials in L1 through the packer instead of reloading DST — NOT yet implemented by the emitted compute kernel, so the emitter refuses it (see ComputeKnobs::check_supported).

pcc_floor: f64

The PCC the emitted program requires of itself to print LOGOS_MM_PASS.

Implementations

impl ComputeKnobs

pub fn from_env() -> Result<Self, String>

The knobs the measurement harness asks for: LOGOS_TT_FIDELITY (LoFi|HiFi2|HiFi3|HiFi4), LOGOS_TT_FP32 (0|1), LOGOS_TT_L1ACC (0|1); an unset variable means the default.

pub fn from_vars( fidelity: Option<&str>, fp32: Option<&str>, l1acc: Option<&str>, ) -> Result<Self, String>

The pure core of ComputeKnobs::from_env.

pub fn dst_tiles(&self, chip: &ChipKnobs) -> usize

The DST tile budget these numerics leave on chip: fp32 accumulate halves it.

pub fn check_supported(&self) -> Result<(), String>

Refuse, typed, any knob the emitted kernels do not actually implement — a program must never print a numerics knob it does not run.

pub fn compute_config_fields(&self) -> String

The .math_fidelity = …, .fp32_dest_acc_en = …, fragment of a tt-metal ComputeConfig.

pub fn pass_tokens(&self) -> String

The knob tokens of the LOGOS_MM_PASS line (the ledger re-parses exactly these).

Trait Implementations

impl Clone for ComputeKnobs

fn clone(&self) -> ComputeKnobs

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ComputeKnobs

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

Formats the value using the given formatter.

impl Default for ComputeKnobs

fn default() -> Self

ttnn’s own bf16 default numerics (HiFi2, bf16 dest-accumulate), so a head-to-head is matched by construction. packer_l1_acc stays off until the kernel implements it.

impl PartialEq for ComputeKnobs

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

impl StructuralPartialEq for ComputeKnobs