Module memory 

Device-memory types.

• DeviceBuffer<T> — owned, typed GPU allocation.
• DeviceSlice<'_, T> / DeviceSliceMut<'_, T> — non-owning views into a buffer, with borrow-checker-tracked lifetimes.

Structs

DeviceBuffer

Owned, typed allocation of device memory.

DeviceSlice

Immutable view into a range of a DeviceBuffer.

DeviceSliceMut

Mutable view into a range of a DeviceBuffer.

ManagedBuffer

Owned allocation of unified (managed) memory — a single pointer that is accessible from both the host and the GPU, with on-demand migration handled by the driver. Compare with DeviceBuffer, which is device-only and requires explicit memcpys.

Enums

ManagedAttach

Attach-mode for ManagedBuffer::new_with_flags.

MemAdvise

Memory-usage advice for cuMemAdvise.

PrefetchTarget

Destination for mem_prefetch_v2 / mem_advise_v2. Composes the CUmemLocation::{type_, id} pair.

Traits

DevicePtr

Anything that can be read as a [T] on the device.

DevicePtrMut

A DevicePtr that supports writes.

Functions

allocation_properties_from_handle

Query the creation props of an existing allocation handle.

get_handle_for_address_range^⚠

Export an OS-level handle (e.g. a DMA-buf file descriptor on Linux) for a size-byte VA range starting at dptr.

mem_advise_v2

cuMemAdvise_v2 — unified-memory hint at a specific location.

mem_get_info

Current device’s free and total global memory, in bytes.

mem_prefetch_v2

cuMemPrefetchAsync_v2 — prefetch count bytes starting at dptr to the given PrefetchTarget, ordered on stream.

memcpy^⚠

Generic byte-count copy that works on any pair of CUDA-addressable pointers (device, unified, host-pinned). Use this when the kind of memory at each end isn’t known at the call site (typical for runtime-decided unified-memory paths).

memcpy_async^⚠

Async variant of memcpy ordered on stream.

memcpy_peer

Peer-to-peer device memory copy between two contexts — the pointers must be valid device pointers in their respective contexts, and peer access must be enabled (see Context::enable_peer_access).

memcpy_peer_async

Async peer-to-peer device memory copy ordered on stream.

memset_2d_u8

2D pitched 8-bit memset: fill a width × height rectangle of bytes at dst (row pitch pitch) with value.

memset_2d_u16

2D pitched 16-bit memset.

memset_2d_u32

2D pitched 32-bit memset.

memset_u8_async

Async 8-bit memset on stream.

memset_u16

Fill count 16-bit elements at dst with value (synchronous).

memset_u32

Synchronous 32-bit memset.

memset_u16_async

Async variant of memset_u16 ordered on stream.

memset_u32_async

Async 32-bit memset on stream.

retain_allocation_handle

Reverse lookup: given a device pointer inside a VMM mapping, bump the underlying allocation handle’s refcount and return it. Pair with cuMemRelease to drop the extra ref.