Struct rust_cuda::host::Stream

source ·

pub struct Stream<'stream> { /* private fields */ }

Available on crate feature host only.

Implementations§

source §

impl<'stream> Stream<'stream>

source

pub fn with<O>( stream: &mut Stream, inner: impl for<'new_stream> FnOnce(Stream<'new_stream>) -> O, ) -> O

Create a new uniquely branded Stream, which can bind async operations to the Stream that they are computed on.

The uniqueness guarantees are provided by using branded types, as inspired by the Ghost Cell paper by Yanovski, J., Dang, H.-H., Jung, R., and Dreyer, D.: https://doi.org/10.1145/3473597.

§Examples

The following example shows that two Stream’s with different 'stream lifetime brands cannot be used interchangeably.

use rust_cuda::host::Stream;

fn check_same<'stream>(_stream_a: Stream<'stream>, _stream_b: Stream<'stream>) {}

fn two_streams<'stream_a, 'stream_b>(stream_a: Stream<'stream_a>, stream_b: Stream<'stream_b>) {
    check_same(stream_a, stream_b);
}

Methods from Deref<Target = Stream>§

source

pub fn get_flags(&self) -> Result<StreamFlags, CudaError>

Return the flags which were used to create this stream.

§Examples

use rustacuda::stream::{Stream, StreamFlags};

let stream = Stream::new(StreamFlags::NON_BLOCKING, None)?;
assert_eq!(StreamFlags::NON_BLOCKING, stream.get_flags().unwrap());

source

pub fn get_priority(&self) -> Result<i32, CudaError>

Return the priority of this stream.

If this stream was created without a priority, returns the default priority. If the stream was created with a priority outside the valid range, returns the clamped priority.

§Examples

use rustacuda::stream::{Stream, StreamFlags};

let stream = Stream::new(StreamFlags::NON_BLOCKING, 1i32.into())?;
println!("{}", stream.get_priority()?);

source

pub fn add_callback<T>(&self, callback: Box<T>) -> Result<(), CudaError>
where T: FnOnce(Result<(), CudaError>) + Send,

Add a callback to a stream.

The callback will be executed after all previously queued items in the stream have been completed. Subsequently queued items will not execute until the callback is finished.

Callbacks must not make any CUDA API calls.

The callback will be passed a CudaResult<()> indicating the current state of the device with Ok(()) denoting normal operation.

§Examples

use rustacuda::stream::{Stream, StreamFlags};

let stream = Stream::new(StreamFlags::NON_BLOCKING, 1i32.into())?;

// ... queue up some work on the stream

stream.add_callback(Box::new(|status| {
    println!("Device status is {:?}", status);
}));

// ... queue up some more work on the stream

source

pub fn synchronize(&self) -> Result<(), CudaError>

Wait until a stream’s tasks are completed.

Waits until the device has completed all operations scheduled for this stream.

§Examples

use rustacuda::stream::{Stream, StreamFlags};

let stream = Stream::new(StreamFlags::NON_BLOCKING, 1i32.into())?;

// ... queue up some work on the stream

// Wait for the work to be completed.
stream.synchronize()?;

source

pub fn wait_event( &self, event: &Event, flags: StreamWaitEventFlags, ) -> Result<(), CudaError>

Make the stream wait on an event.

All future work submitted to the stream will wait for the event to complete. Synchronization is performed on the device, if possible. The event may originate from different context or device than the stream.

§Example

use rustacuda::stream::{Stream, StreamFlags, StreamWaitEventFlags};
use rustacuda::event::{Event, EventFlags};

let stream_0 = Stream::new(StreamFlags::NON_BLOCKING, None)?;
let stream_1 = Stream::new(StreamFlags::NON_BLOCKING, None)?;
let event = Event::new(EventFlags::DEFAULT)?;

// do some work on stream_0 ...

// record an event
event.record(&stream_0)?;

// wait until the work on stream_0 is finished before continuing stream_1
stream_1.wait_event(event, StreamWaitEventFlags::DEFAULT)?;
}