Struct TensorDesc 

pub struct TensorDesc<T: GpuFloat> {
    pub ptr: CUdeviceptr,
    pub dims: Vec<u32>,
    pub strides: Vec<u32>,
    pub layout: TensorLayout,
    /* private fields */
}

Immutable tensor descriptor binding a device pointer to shape metadata.

TensorDesc does not own the device memory; it merely borrows the raw pointer for the duration of a DNN operation. The caller must ensure that the referenced DeviceBuffer outlives any computation that uses this descriptor.

Fields

ptr: CUdeviceptr

Raw device pointer to the first element.

dims: Vec<u32>

Shape (one entry per dimension).

strides: Vec<u32>

Strides (one entry per dimension, in elements not bytes).

layout: TensorLayout

Memory layout convention.

Implementations

impl<T: GpuFloat> TensorDesc<T>

pub fn nchw( buf: &DeviceBuffer<T>, n: u32, c: u32, h: u32, w: u32, ) -> DnnResult<Self>

Creates an NCHW tensor descriptor from a device buffer.

Errors

Returns DnnError::InvalidDimension if any dimension is zero. Returns DnnError::BufferTooSmall if the buffer cannot hold n * c * h * w elements.

pub fn nhwc( buf: &DeviceBuffer<T>, n: u32, c: u32, h: u32, w: u32, ) -> DnnResult<Self>

Creates an NHWC tensor descriptor from a device buffer.

Errors

Same as nchw.

pub fn ncdhw( buf: &DeviceBuffer<T>, n: u32, c: u32, d: u32, h: u32, w: u32, ) -> DnnResult<Self>

Creates an NCDHW (3-D volumetric) tensor descriptor.

Errors

Same as nchw.

pub fn matrix(buf: &DeviceBuffer<T>, rows: u32, cols: u32) -> DnnResult<Self>

Creates a 2-D matrix descriptor (rows x cols, row-major).

Errors

Returns DnnError::InvalidDimension if either dimension is zero. Returns DnnError::BufferTooSmall if the buffer is too small.

pub fn from_raw( ptr: CUdeviceptr, dims: Vec<u32>, strides: Vec<u32>, layout: TensorLayout, ) -> DnnResult<Self>

Constructs a descriptor from raw components without buffer validation.

The caller must ensure that ptr points to a valid device allocation large enough for the described tensor.

pub fn numel(&self) -> usize

Returns the total number of elements in the tensor.

pub fn ndim(&self) -> usize

Returns the number of dimensions.

pub fn validate_buffer_size(&self, buf: &DeviceBuffer<T>) -> DnnResult<()>

Validates that buf is large enough to hold this tensor.

Errors

Returns DnnError::BufferTooSmall if the buffer has fewer elements than numel.