Struct GpuBuffer 

pub struct GpuBuffer<T> { /* private fields */ }

A typed buffer in GPU device memory, wrapping cudarc’s CudaSlice<T>.

Created via KaioDevice::alloc_from or KaioDevice::alloc_zeros.

Memory management

GpuBuffer does not implement Drop manually — cudarc’s CudaSlice handles device memory deallocation automatically when the buffer is dropped. The CudaSlice holds an Arc<CudaContext> internally, ensuring the CUDA context outlives the allocation.

Implementations

impl<T> GpuBuffer<T>

pub fn len(&self) -> usize

Number of elements in the buffer.

pub fn is_empty(&self) -> bool

Returns true if the buffer contains no elements.

pub fn inner(&self) -> &CudaSlice<T>

Access the underlying CudaSlice for passing to cudarc launch operations.

This is the escape hatch for Sprint 1.7’s launch builder — the caller pushes &buf.inner() as a kernel argument.

pub fn inner_mut(&mut self) -> &mut CudaSlice<T>

Mutable access to the underlying CudaSlice.

impl<T: DeviceRepr + Default + Clone + Unpin> GpuBuffer<T>

pub fn to_host(&self, device: &KaioDevice) -> Result<Vec<T>>

Transfer buffer contents from device to host.

Requires a reference to the KaioDevice that created this buffer (for stream access). The device is borrowed, not consumed.

Example

let device = KaioDevice::new(0)?;
let buf = device.alloc_from(&[1.0f32, 2.0, 3.0])?;
let host_data = buf.to_host(&device)?;
assert_eq!(host_data, vec![1.0, 2.0, 3.0]);