Struct DataBuffer 

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

Data storage for tensor elements.

Abstracts over ownership: data may be Rust-owned (Vec<T>), externally-owned (e.g., imported via DLPack with a release callback), or GPU device memory. Shape and stride metadata are not stored here.

Shared ownership

DataBuffer uses Arc, so cloning a tensor or buffer is shallow and shares storage. Deep copies happen only when a tensor is materialized into a new contiguous allocation.

Examples

use tenferro_tensor::DataBuffer;

let buf = DataBuffer::<f64>::from_vec(vec![1.0, 2.0, 3.0]);
assert_eq!(buf.len(), 3);

Implementations

impl<T> DataBuffer<T>

pub fn from_vec(v: Vec<T>) -> Self

Create a buffer from an owned Vec<T>.

Examples

use tenferro_tensor::DataBuffer;

let buf = DataBuffer::from_vec(vec![1.0, 2.0, 3.0]);
assert!(buf.is_owned());

pub unsafe fn from_external( ptr: *const T, len: usize, release: impl FnOnce() + Send + 'static, ) -> Self

Create a buffer from externally-owned data with a release callback.

Safety

• ptr must point to a valid, properly aligned allocation of at least len elements of T.
• The allocation must remain valid until the release callback fires.

Examples

use tenferro_tensor::DataBuffer;

let data = vec![1.0, 2.0, 3.0];
let ptr = data.as_ptr();
let len = data.len();
let buf = unsafe { DataBuffer::from_external(ptr, len, move || drop(data)) };
assert!(!buf.is_owned());

pub fn as_slice(&self) -> Option<&[T]>

Returns the raw data as a slice for CPU-accessible buffers.

Examples

use tenferro_tensor::DataBuffer;

let buf = DataBuffer::from_vec(vec![1.0, 2.0, 3.0]);
assert_eq!(buf.as_slice(), Some(&[1.0, 2.0, 3.0][..]));

pub fn as_mut_slice(&mut self) -> Option<&mut [T]>

Returns the raw data as a mutable slice, if uniquely owned.

Examples

use tenferro_tensor::DataBuffer;

let mut buf = DataBuffer::from_vec(vec![1.0, 2.0]);
if let Some(slice) = buf.as_mut_slice() {
    slice[0] = 42.0;
}
assert_eq!(buf.as_slice().unwrap()[0], 42.0);

pub fn len(&self) -> usize

Returns the number of elements in the buffer.

Examples

use tenferro_tensor::DataBuffer;

let buf = DataBuffer::from_vec(vec![1.0, 2.0, 3.0]);
assert_eq!(buf.len(), 3);

pub fn is_empty(&self) -> bool

Returns true if the buffer has no elements.

Examples

use tenferro_tensor::DataBuffer;

let buf = DataBuffer::<f64>::from_vec(vec![]);
assert!(buf.is_empty());

pub fn is_owned(&self) -> bool

Returns true if the buffer is Rust-owned.

Examples

use tenferro_tensor::DataBuffer;

let buf = DataBuffer::from_vec(vec![1.0f64]);
assert!(buf.is_owned());

pub fn is_gpu(&self) -> bool

Returns true if the buffer resides on GPU device memory.

Examples

use tenferro_tensor::DataBuffer;

let buf = DataBuffer::from_vec(vec![1.0f64]);
assert!(!buf.is_gpu());

pub fn is_unique(&self) -> bool

Returns true if this is the only reference to the underlying buffer.

Examples

use tenferro_tensor::DataBuffer;

let buf = DataBuffer::from_vec(vec![1.0f64]);
assert!(buf.is_unique());
let buf2 = buf.clone();
assert!(!buf.is_unique());

pub fn try_into_vec(self) -> Option<Vec<T>>

Extract the inner Vec<T> if this is the sole owner of a CPU-owned buffer.

pub fn as_ptr(&self) -> Option<*const T>

Returns a raw CPU pointer to the data, or None for GPU buffers.

Examples

use tenferro_tensor::DataBuffer;

let buf = DataBuffer::from_vec(vec![1.0f64]);
assert!(buf.as_ptr().is_some());

pub fn as_device_ptr(&self) -> Option<*const T>

Returns the GPU device pointer, or None for CPU buffers.

Examples

use tenferro_tensor::DataBuffer;

let buf = DataBuffer::from_vec(vec![1.0f64]);
assert!(buf.as_device_ptr().is_none());

pub fn gpu_memory_space(&self) -> Option<LogicalMemorySpace>

Returns the logical memory space of a GPU buffer, or None for CPU buffers.

Examples

use tenferro_tensor::DataBuffer;

let buf = DataBuffer::from_vec(vec![1.0f64]);
assert!(buf.gpu_memory_space().is_none());

impl<T: Scalar> DataBuffer<T>

pub fn zeros_on_device( n_elements: usize, memory_space: LogicalMemorySpace, ) -> Result<Self>

Allocate a zero-filled buffer on the specified device.

For CPU (MainMemory) this creates an owned Vec<T> filled with T::zero(). For CUDA GPU memory (when the cuda feature is enabled) this allocates directly on the device and zero-fills via host-to-device copy.

Errors

Returns an error for unsupported memory spaces or if device allocation fails.

Examples

use tenferro_tensor::DataBuffer;
use tenferro_device::LogicalMemorySpace;

let buf = DataBuffer::<f64>::zeros_on_device(4, LogicalMemorySpace::MainMemory).unwrap();
assert_eq!(buf.as_slice().unwrap(), &[0.0; 4]);

pub fn allocate_uninit_on_device( n_elements: usize, memory_space: LogicalMemorySpace, ) -> Result<Self>

Allocate an uninitialized buffer on the specified device.

For CPU (MainMemory) this creates an owned Vec<T> with the requested capacity. The contents are unspecified (currently zero-filled for safety, but callers must not rely on that).

For CUDA GPU memory (when the cuda feature is enabled) this allocates device memory without initialization.

Errors

Returns an error for unsupported memory spaces or if device allocation fails.

Examples

use tenferro_tensor::DataBuffer;
use tenferro_device::LogicalMemorySpace;

let buf = DataBuffer::<f64>::allocate_uninit_on_device(4, LogicalMemorySpace::MainMemory).unwrap();
assert_eq!(buf.len(), 4);

Trait Implementations

impl<T> Clone for DataBuffer<T>

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl<T: Send> Send for DataBuffer<T>

impl<T: Sync> Sync for DataBuffer<T>