Struct TensorMap 

pub struct TensorMap(/* private fields */);

A 128-byte TMA descriptor, aligned as the driver requires.

Implementations

impl TensorMap

pub fn bytes(&self) -> &[u8; 128]

The raw bytes, for KernelLaunch::arg_tensormap.

pub fn tiled_2d_bf16( global: DevicePtr, rows: u64, cols: u64, row_stride_elems: u64, box_rows: u32, box_cols: u32, ) -> Result<Self>

Describe a row-major bf16 matrix as tiles of box_rows x box_cols.

rows/cols and row_stride_elems are in ELEMENTS, in the row-major terms the rest of the codebase uses; the fastest-varying-first ordering the driver wants is applied here so callers never have to think about it. row_stride_elems is the distance between consecutive rows and may exceed cols (a view into a wider tensor).

Loads of tiles that hang off the end are ZERO-FILLED by hardware, so the caller does not predicate the tail.

Trait Implementations

impl Clone for TensorMap

fn clone(&self) -> TensorMap

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TensorMap

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

The 128 bytes are an opaque driver-owned blob; printing them is noise.

impl Copy for TensorMap