Struct NativeMemoryPool 

pub struct NativeMemoryPool { /* private fields */ }

Thin wrapper around the CUDA driver’s stream-ordered memory pool (cuMemPoolCreate / cuMemPoolDestroy).

Allocations are issued via NativeMemoryPool::alloc_async which invokes cuMemAllocFromPoolAsync; frees are issued via NativeMemoryPool::free_async which invokes cuMemFreeAsync.

Stream-ordering

The CUDA stream-ordered pool API requires the caller to ensure all outstanding work on the stream has completed before destroying the pool. The Drop implementation calls cuMemPoolDestroy and silently swallows any error to honour the standard Drop convention. Call NativeMemoryPool::destroy explicitly to surface destruction errors.

Status

On systems without a CUDA driver (e.g. macOS), NativeMemoryPool::new fails with CudaError::NotInitialized. On older drivers that lack the pool entry points it fails with CudaError::NotSupported.

Implementations

impl NativeMemoryPool

pub fn new(props: NativeMemoryPoolProps) -> CudaResult<Self>

Creates a new native memory pool on the device described by props.

Errors

• CudaError::InvalidValue if device_ordinal is negative.
• CudaError::NotInitialized if no CUDA driver is available.
• CudaError::NotSupported if the driver does not export cuMemPoolCreate.
• Other CudaError variants on driver failure.

pub fn raw(&self) -> CUmemoryPool

Returns the raw CUmemoryPool handle.

pub fn device_ordinal(&self) -> i32

Returns the device ordinal that backs this pool.

pub fn alloc_async( &self, bytes: usize, stream: &Stream, ) -> CudaResult<CUdeviceptr>

Asynchronously allocates bytes of memory from the pool, ordered against stream.

Errors

• CudaError::InvalidValue if bytes is zero.
• CudaError::NotInitialized if no CUDA driver is available.
• CudaError::NotSupported if the driver does not export cuMemAllocFromPoolAsync.
• Other CudaError variants on driver failure.

pub fn free_async(&self, ptr: CUdeviceptr, stream: &Stream) -> CudaResult<()>

Asynchronously frees a pointer previously returned by alloc_async, ordered against stream.

Errors

• CudaError::NotInitialized if no CUDA driver is available.
• CudaError::NotSupported if the driver does not export cuMemFreeAsync.
• Other CudaError variants on driver failure.

pub fn destroy(self) -> CudaResult<()>

Destroys the pool, returning any driver error to the caller.

The caller is responsible for ensuring all outstanding work on streams that allocated from this pool has completed before calling destroy.

After this call returns, the Drop implementation will be a no-op.

Errors

• CudaError::NotInitialized if no CUDA driver is available.
• CudaError::NotSupported if the driver does not export cuMemPoolDestroy.
• Other CudaError variants on driver failure.

Trait Implementations

impl Drop for NativeMemoryPool

fn drop(&mut self)

Executes the destructor for this type.

impl Send for NativeMemoryPool

impl Sync for NativeMemoryPool