Struct Executor 

pub struct Executor<'a> { /* private fields */ }

Runner of batches of one model, with its lane pools built once.

Implementations

impl<'a> Executor<'a>

pub fn new( entry: &'a ModelEntry, gpu: Option<&'a GpuContext>, measure: Arc<MeasurePlan>, layout: ExecutionLayout, control: SweepControl, ) -> Result<Self, ExecutionError>

Returns an executor for entry sampled as measure specifies, with a pool for each lane of layout.

A single lane gets its own pool as well, even a lane as wide as rayon’s global pool. Note that a target that cannot spawn threads builds no pool, and runs a CPU model in a single lane on the calling thread, whatever layout requests.

Errors

Returns ExecutionError::NoDevice for a GPU model with no device in gpu, and ExecutionError::Pool when a lane’s pool cannot be built.

pub fn with_timeout(self, timeout: Option<Duration>) -> Self

Returns the executor with each run abandoned once its stepping and sampling pass timeout.

The timeout is checked between slices of steps. A GPU run counts its share of the wall time since its build. The time spent visiting the tracks is split evenly between the live runs, and the time the batch spent paused is left out.

pub fn with_active_runs(self, active_runs: Option<ActiveRuns>) -> Self

Returns the executor with each run listed in active_runs from its build to its end.

pub fn with_gpu_memory_budget(self, budget: Option<u64>) -> Self

Returns the executor with a budget of budget bytes of device memory for the live GPU runs, as gpu_memory_budget interprets it.

A run is built once its demand fits the budget beside the demand of the live runs. A run larger than the budget is built once no other run is live, and runs alone.

pub fn with_track_depth(self, depth: GpuTrackDepth) -> Self

Returns the executor with each GPU track keeping depth on the device.

Panics

Panics when depth holds no command buffer or no step, or more steps than one submission can hold.

pub fn layout(&self) -> ExecutionLayout

Layout the executor runs on, after Self::new adjusted it to the target.

pub fn control(&self) -> &SweepControl

Control the executor reads between slices of steps.

pub fn run_batch( &self, requests: &[RunRequest<'_>], sink: &mut dyn RunSink, ) -> Result<BatchEnd, ExecutionError>

Runs every request and commits each outcome to sink in request order.

Runs that finish early wait in memory until every earlier request is committed. Note that on wasm32 a pause panics on a thread that cannot wait, as SweepControl::proceed describes.

Errors

Returns ExecutionError when a lane’s thread cannot start or panics outside a run, or sink rejects a run.

Trait Implementations

impl<'a> Debug for Executor<'a>

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

Formats the value using the given formatter.