Struct StreamPool 

pub struct StreamPool { /* private fields */ }

Pool of owned non-blocking CUDA streams.

Implementations

impl StreamPool

pub fn new(device: Arc<CudaDevice>, max_streams: usize) -> Self

Construct a pool bound to device, capped at max_streams.

pub fn with_defaults(device: Arc<CudaDevice>) -> Self

Construct with the default cap DEFAULT_MAX_STREAMS.

pub fn acquire(&self) -> Result<StreamId, StreamPoolError>

Acquire a non-default stream id, growing the pool up to max_streams. Each successful call returns a distinct StreamId backed by an owned non-blocking cudarc::driver::CudaStream forked from the device’s default stream.

Errors

• StreamPoolError::Capacity if the pool already holds max_streams non-default streams.
• StreamPoolError::ForkFailed if the underlying CudaStream::fork call failed.

Streams are never returned to a free-list; they remain valid for the runtime’s lifetime so previously returned StreamId handles keep resolving.

pub fn resolve(&self, id: StreamId) -> Option<Arc<CudaStream>>

Borrow the live CudaStream for id. Returns None if id has never been issued by this pool. The default-stream slot resolves to the device’s default stream.

pub fn non_default_len(&self) -> usize

Number of non-default streams currently in the pool.

pub fn device(&self) -> &Arc<CudaDevice>

Borrow the device handle. Test helpers use this to launch kernels into the same device the pool was constructed on.

pub fn max_streams(&self) -> usize

Maximum streams the pool will create on demand.

pub fn pool_bytes_per_stream(&self) -> u64

Planned per-stream pool budget, in bytes, from XLOG_WCOJ_POOL_MB_PER_STREAM or the 256 MiB default.