Struct TensorMapDescriptor 

pub struct TensorMapDescriptor {
    pub data_type: TensorMapDataType,
    pub global_address: usize,
    pub global_dim: Vec<u64>,
    pub global_strides: Vec<u64>,
    pub box_dim: Vec<u32>,
    pub element_strides: Vec<u32>,
    pub interleave: TensorMapInterleave,
    pub swizzle: TensorMapSwizzle,
    pub l2_promotion: TensorMapL2Promotion,
    pub oob_fill: TensorMapOobFill,
}

Host-side builder for the tiled flavour of cuTensorMapEncodeTiled.

Cap at 5 dimensions (the CUDA-driver hard limit). Validation:

• rank must equal global_dim.len() and box_dim.len() and element_strides.len(); global_strides.len() == rank - 1.
• global_address must be 16-byte aligned (TMA requires it).
• Every entry of global_dim and box_dim must be non-zero.

Fields

data_type: TensorMapDataType

global_address: usize

global_dim: Vec<u64>

global_strides: Vec<u64>

box_dim: Vec<u32>

element_strides: Vec<u32>

interleave: TensorMapInterleave

swizzle: TensorMapSwizzle

l2_promotion: TensorMapL2Promotion

oob_fill: TensorMapOobFill

Implementations

impl TensorMapDescriptor

pub fn new(data_type: TensorMapDataType, global_address: usize) -> Self

Construct a defaulted-shape descriptor. Caller must populate global_dim, global_strides, box_dim, element_strides before calling TensorMapDescriptor::validate.

pub fn rank(&self) -> usize

Tensor rank (1..=5).

pub fn validate(&self) -> Result<(), TmaEncodeError>

Validate that all sizes line up and the global address is 16B aligned. Returns Err(TmaEncodeError::*) on mismatch. Pure host validation — does not call into the driver.

Trait Implementations

impl Clone for TensorMapDescriptor

fn clone(&self) -> TensorMapDescriptor

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TensorMapDescriptor

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.