Struct TunableSet 

pub struct TunableSet<K, F, Output>
where
    K: AutotuneKey,
    F: TuneInputs,
    Output: 'static,
{ /* private fields */ }

A set of candidate tunable functions for autotune, sharing a key generator and an input generator. See TuneInputs for the F parameter.

Implementations

impl<K, F, Output> TunableSet<K, F, Output>

where K: AutotuneKey, F: TuneInputs, Output: 'static,

pub fn len(&self) -> usize

The number of tunables in the set.

pub fn is_empty(&self) -> bool

Whether this set contains no tunables.

pub fn new( key_gen: impl KeyGenerator<K, F>, input_gen: impl InputGenerator<K, F>, ) -> TunableSet<K, F, Output>

Create a tunable set from a key generator and an input generator.

pub fn with_stack_tuning( self, reference: usize, revision: &str, workload: impl for<'a> Fn(&<F as TuneInputs>::At<'a>) -> String + Send + Sync + 'static, ) -> TunableSet<K, F, Output>

Register an exact workload signature and a known-safe reference for full-stack tuning. The input generator MUST isolate all mutable state for each benchmark invocation. Include dtype, exact shape/strides, precision, optional inputs and semantic parameters.

pub fn stack_reference(&self) -> Option<usize>

pub fn stack_workload<'a>( &self, inputs: &<F as TuneInputs>::At<'a>, ) -> Option<String>

pub fn new_cloning_inputs( key_gen: impl KeyGenerator<K, F>, ) -> TunableSet<K, F, Output>

Shorthand for new with a [CloneInputGenerator]: benchmarks run on clones of the real call inputs.

pub fn with(self, tunable: Tunable<K, F, Output>) -> TunableSet<K, F, Output>

Register a tunable with this tunable set.

pub fn autotunables(&self) -> impl Iterator<Item = &TuneFn<F, Output>>

All candidate operations in this set, in registration order.

pub fn fastest(&self, fastest_index: usize) -> &TuneFn<F, Output>

Returns the operation for the given index, matching the order returned by autotunables. Tunables are tried in order, so index 0 should be a good default.

pub fn compute_checksum(&self) -> String

Compute a checksum that invalidates outdated cached auto-tune results when the set of tunable names changes.

pub fn stack_checksum(&self) -> String

Separate, length-delimited manifest for the opt-in full-stack cache.

pub fn generate_key<'a>(&self, inputs: &<F as TuneInputs>::At<'a>) -> K

Generate a key from a set of inputs

pub fn generate_inputs<'a>( &self, key: &K, inputs: &<F as TuneInputs>::At<'a>, ) -> <F as TuneInputs>::At<'a>

Generate a set of test inputs from a key and reference inputs.