Struct PooledBuffer 

pub struct PooledBuffer<T: Copy> { /* private fields */ }

A device buffer allocated from a MemoryPool.

Unlike DeviceBuffer, a PooledBuffer is freed asynchronously — the free operation is enqueued on the stream rather than blocking the CPU. This enables overlap of allocation, computation, and deallocation across multiple stream operations.

Stream-ordering

Dropping a PooledBuffer does not immediately reuse its device pointer. Instead, a CU_EVENT_DISABLE_TIMING event is recorded on the stream the buffer was allocated on (the stream argument to alloc_async); the pointer is only handed back out to a later alloc_async call once that event has completed, i.e. once all work enqueued on the stream before the drop has actually finished executing on the device. A PooledBuffer must therefore not outlive its stream. If the recycle event cannot be created or recorded, Drop falls back to a blocking cuStreamSynchronize on the owning stream before freeing the pointer directly, so correctness is preserved at the cost of losing the pool’s overlap benefits for that allocation.

Status

This type allocates from an in-process memory pool and returns buffers to that pool on drop.

Implementations

impl<T: Copy> PooledBuffer<T>

pub fn alloc_async( pool: &MemoryPool, n: usize, stream: &Stream, ) -> CudaResult<Self>

Asynchronously allocates a buffer of n elements from the given pool.

The allocation is ordered relative to other operations on stream; stream is also recorded against so that Drop can establish that all work using this buffer has completed before its pointer is recycled to a later caller (see the type-level docs).

Errors

• CudaError::InvalidValue if n is zero or n * size_of::<T>() overflows.
• Other CudaError variants if the underlying allocation fails.

pub fn len(&self) -> usize

Returns the number of T elements in this buffer.

pub fn is_empty(&self) -> bool

Returns true if the buffer contains zero elements.

pub fn byte_size(&self) -> usize

Returns the total size of the allocation in bytes.

pub fn as_device_ptr(&self) -> CUdeviceptr

Returns the raw CUdeviceptr handle.

Trait Implementations

impl<T: Copy> Drop for PooledBuffer<T>

fn drop(&mut self)

Executes the destructor for this type.

fn pin_drop(self: Pin<&mut Self>)

🔬This is a nightly-only experimental API. (pin_ergonomics)

Execute the destructor for this type, but different to Drop::drop, it requires self to be pinned.