Module unified 

Unified (managed) memory buffer.

UnifiedBuffer<T> wraps cuMemAllocManaged, which allocates memory that is automatically migrated between host and device by the CUDA Unified Memory subsystem. The allocation is accessible from both CPU code (via as_slice / as_mut_slice) and GPU kernels (via as_device_ptr).

Coherence caveat

The host-side accessors are only safe to call when no GPU kernel is concurrently reading or writing the same memory. After launching a kernel that touches a unified buffer, synchronise the stream (or the entire context) before accessing the data from the host.

Ownership

The allocation is freed with cuMemFree_v2 on drop. Errors during drop are logged via tracing::warn.

Example

let mut ubuf = UnifiedBuffer::<f32>::alloc(512)?;
// Write from the host side (no kernel running).
for (i, v) in ubuf.as_mut_slice().iter_mut().enumerate() {
    *v = i as f32;
}
// Pass ubuf.as_device_ptr() to a kernel…

Structs

UnifiedBuffer

A contiguous buffer of T elements in CUDA unified (managed) memory.