Module copy_2d3d 

2D and 3D memory copy operations for pitched and volumetric data.

GPU memory is often allocated as “pitched” 2D arrays where each row has padding bytes to satisfy alignment requirements. The standard 1D copy functions cannot handle this row padding — they would copy the padding bytes as if they were data.

This module provides:

• Memcpy2DParams — parameters for 2D (row-padded) copies.
• Memcpy3DParams — parameters for 3D (volumetric, doubly-padded) copies.
• Copy functions for host-to-device, device-to-host, and device-to-device transfers in 2D and 3D.

Pitch vs Width

• pitch — total bytes per row including alignment padding.
• width — bytes of actual data per row to copy.

The pitch must be >= width for both source and destination.

Status

The CUDA driver function cuMemcpy2D_v2 is now wired through oxicuda-driver. All 2D copy variants (DtoD, HtoD, DtoH) build a CUDA_MEMCPY2D descriptor and invoke the driver entry point when available. cuMemcpy3D_v2 is not yet loaded; the 3D copy still returns CudaError::NotSupported when a driver is present but the symbol is missing.

Example

use oxicuda_memory::copy_2d3d::{Memcpy2DParams, copy_2d_dtod};
use oxicuda_memory::DeviceBuffer;

let params = Memcpy2DParams {
    src_pitch: 512,
    dst_pitch: 512,
    width: 480,      // 480 bytes of data per row
    height: 256,     // 256 rows
};

let mut dst = DeviceBuffer::<u8>::alloc(512 * 256)?;
let src = DeviceBuffer::<u8>::alloc(512 * 256)?;
copy_2d_dtod(&mut dst, &src, &params)?;

Structs

Memcpy2DParams

Parameters for a 2D (pitched) memory copy.

Memcpy3DParams

Parameters for a 3D (volumetric) memory copy.

Functions

copy_2d_dtod

Copies a 2D region between two device buffers (device-to-device).

copy_2d_dtoh

Copies a 2D region from a device buffer to host memory.

copy_2d_htod

Copies a 2D region from host memory to a device buffer.

copy_3d_dtod

Copies a 3D region between two device buffers (device-to-device).