Module peer_copy 

Peer-to-peer (P2P) memory copy operations for multi-GPU workloads.

This module provides functions to check, enable, and disable peer access between CUDA devices, as well as copy data between device buffers on different GPUs.

Peer access enables direct GPU-to-GPU memory transfers over PCIe or NVLink without staging through host memory, significantly improving transfer bandwidth in multi-GPU configurations.

Platform note

The underlying cuDeviceCanAccessPeer, cuCtxEnablePeerAccess, cuCtxDisablePeerAccess, and cuMemcpyPeer driver functions are not yet loaded by oxicuda-driver. All functions currently return CudaError::NotSupported as placeholders. The API surface is established here so that downstream crates can program against it.

Example

use oxicuda_driver::device::Device;
use oxicuda_memory::peer_copy;

oxicuda_driver::init()?;
let dev0 = Device::get(0)?;
let dev1 = Device::get(1)?;

if peer_copy::can_access_peer(&dev0, &dev1)? {
    peer_copy::enable_peer_access(&dev0, &dev1)?;
    // Now D2D copies between dev0 and dev1 can go over NVLink/PCIe
    // peer_copy::copy_peer(&mut dst_buf, &dev1, &src_buf, &dev0)?;
}

Functions

can_access_peer

Checks whether device can directly access memory on peer.

copy_peer

Copies data between device buffers on different GPUs (synchronous).

copy_peer_async

Copies data between device buffers on different GPUs (asynchronous).

disable_peer_access

Disables peer access from the current context’s device to peer.

enable_peer_access

Enables peer access from the current context’s device to peer.