Struct ThroughputBenchmarker 

pub struct ThroughputBenchmarker { /* private fields */ }

A marker for measuring throughput of compute kernels.

Implementations

impl ThroughputBenchmarker

pub fn new(cache: Arc<Mutex<ThroughputCache>>) -> Self

Creates a new ThroughputBenchmarker with the given cache.

pub fn measure( &mut self, key: ThroughputKey, probe: impl FnOnce() -> Result<ThroughputValue, ThroughputError>, ) -> Result<ThroughputValue, ThroughputError>

The value for key, measured by probe unless the cache holds one.

Errors

Whatever probe reports. Only a measurement is cached.

pub fn sample(kernel_config: KernelConfig) -> ThroughputValue

Warm one shape of a kernel up to its plateau, then keep its fastest sample.

pub fn warm(kernel_config: &KernelConfig) -> usize

Warms the device on one shape and reports the count a sample should carry. It is the device that is warmed, not the shape, so a sweep pays once.

pub fn sample_at( kernel_config: &KernelConfig, iterations: usize, ) -> ThroughputValue

Keeps the fastest sample of a shape the device is already warm on, at the count warm settled.

pub fn rank(kernel_config: &KernelConfig, iterations: usize) -> Ranked

Times one shape briefly, to order it against the others rather than to report its peak, and reports the count the next shape starts from.