Struct TensorMap 

pub struct TensorMap { /* private fields */ }

A 128-byte Hopper TMA descriptor. Pass to a kernel as a __grid_constant__ parameter of type CUtensorMap for use with TMA instructions.

Implementations

impl TensorMap

pub fn encode_tiled( data_type: i32, global_base: CUdeviceptr, global_dim: &[u64], global_strides: &[u64], box_dim: &[u32], element_strides: &[u32], interleave: i32, swizzle: i32, l2_promotion: i32, oob_fill: i32, ) -> Result<Self>

Build a tiled TMA descriptor.

• data_type: element type (one of the DataType::* constants).
• global_base: pointer to the first element of the tensor.
• global_dim: per-axis size of the global tensor (innermost-to-outermost).
• global_strides: per-axis byte strides between successive elements.
• box_dim: per-axis shape of the tile copied at a time.
• element_strides: per-axis element-strides (typically all 1).

All arrays must have length rank = global_dim.len().

pub fn replace_address(&mut self, new_base: CUdeviceptr) -> Result<()>

Swap the global base address of an existing descriptor in place. Lets you reuse one TensorMap across multiple buffers of the same shape/stride.

pub fn as_raw(&self) -> &CUtensorMap

Raw pointer to the 128-byte descriptor — pass this to kernels that take a CUtensorMap parameter.

pub fn as_raw_mut(&mut self) -> &mut CUtensorMap

Mutable raw access (for FFI calls that want *mut CUtensorMap).

Trait Implementations

impl Debug for TensorMap

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

Formats the value using the given formatter.