Module memory 

Runtime-API device memory.

Modules

pinned_flags

Flags for cudaHostAlloc. See baracuda_cuda_sys::runtime::types::cudaHostAllocFlags for raw values.

Structs

DeviceBuffer

Owned, typed allocation of device memory (Runtime API).

ManagedBuffer

Unified managed-memory buffer — allocated via cudaMallocManaged. Accessible from both host and device without explicit copies.

PinnedHostBuffer

Pinned (page-locked) host allocation — CUDA-owned memory that supports real async H↔D copies without staging.

PinnedRegistration

RAII guard for cudaHostRegister — pins an existing host slice and unregisters on drop.

Enums

PrefetchTarget

Target for mem_prefetch_async / mem_advise. The CUDA Runtime API’s v1 variants take an ordinal — pass cudaCpuDeviceId (-1) for host.

Functions

mem_advise^⚠

cudaMemAdvise — unified-memory placement hint. advice is a constant from baracuda_cuda_sys::runtime::types::cudaMemoryAdvise.

mem_get_info

cudaMemGetInfo — (free, total) bytes on the current device.

mem_prefetch_async^⚠

Prefetch count bytes of unified memory at dev_ptr to target, ordered on stream. dev_ptr must be a managed-memory allocation (from ManagedBuffer or cudaMallocManaged).

memcpy_peer

Peer-to-peer device memory copy. Both buffers must be on enabled-peer devices (see crate::Device::enable_peer_access).

memcpy_peer_async

Async peer-to-peer memcpy ordered on stream.