Struct NaturalGradientPolicy 

pub struct NaturalGradientPolicy {
    pub temperature_ns: u64,
}

Natural-gradient policy for choosing the next autotune probe from measured latency samples.

The policy treats the candidate set as a discrete distribution over launch configurations. Latency samples become a softmax over -elapsed_ns / temperature_ns; the supplied inverse-Fisher square-root matrix preconditions that probability/gradient vector before the driver picks the next candidate. CUDA/self-substrate can produce the same fixed-point matrix through the primitive-backed natural-gradient path.

Fields

temperature_ns: u64

Softmax temperature in nanoseconds. Larger values explore more.

Implementations

impl NaturalGradientPolicy

pub fn suggest( &self, measurements: &[TuningMeasurement], fisher_inv_sqrt_q16: &[u32], ) -> Result<NaturalGradientTuningStep, NaturalGradientTuningError>

Suggest the next workgroup-size candidate from latency samples and an inverse-Fisher square-root matrix.

fisher_inv_sqrt_q16 is row-major n x n, 16.16 fixed-point. Passing an identity matrix makes the policy reduce to the softmax-gradient candidate. Non-identity blocks let the runtime bias exploration by the local latency manifold instead of blindly reusing the single fastest point.

Errors

Returns NaturalGradientTuningError when the measurement set is empty, temperature is zero, or the Fisher matrix shape does not match.

Trait Implementations

impl Clone for NaturalGradientPolicy

fn clone(&self) -> NaturalGradientPolicy

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for NaturalGradientPolicy

impl Debug for NaturalGradientPolicy

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

Formats the value using the given formatter.

impl Default for NaturalGradientPolicy

fn default() -> Self

Returns the “default value” for a type.

impl Eq for NaturalGradientPolicy

impl PartialEq for NaturalGradientPolicy

fn eq(&self, other: &NaturalGradientPolicy) -> 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 StructuralPartialEq for NaturalGradientPolicy