Struct TmaEncodeTiledParams 

pub struct TmaEncodeTiledParams {
    pub data_type: CuTensorMapDataType,
    pub num_dims: u32,
    pub global_dims: [u64; 5],
    pub global_strides: [u64; 4],
    pub box_dims: [u32; 5],
    pub element_strides: [u32; 5],
    pub interleave: CuTensorMapInterleave,
    pub swizzle: CuTensorMapSwizzle,
    pub l2_promotion: CuTensorMapL2Promotion,
    pub oob_fill: CuTensorMapFloatOobFill,
}

All parameters required by cuTensorMapEncodeTiled.

Produced by TmaDescriptorBuilder. Pass these to the driver function pointer available in DriverApi::cu_tensor_map_encode_tiled.

Dimension ordering

CUDA TMA uses column-major ordering for dimensions: global_dims[0] is the innermost (fastest-varying) dimension. For a row-major matrix of shape R × C, set global_dims[0] = C (cols) and global_dims[1] = R (rows).

Fields

data_type: CuTensorMapDataType

Element data type.

num_dims: u32

Number of tensor dimensions (1–5).

global_dims: [u64; 5]

Size of each global tensor dimension (innermost first).

global_strides: [u64; 4]

Byte stride between elements in outer dimensions (innermost stride omitted).

box_dims: [u32; 5]

Size of each tile dimension (must fit in shared memory).

element_strides: [u32; 5]

Element stride within each tile dimension (typically all-ones).

interleave: CuTensorMapInterleave

Interleave mode.

swizzle: CuTensorMapSwizzle

Swizzle mode.

l2_promotion: CuTensorMapL2Promotion

L2 promotion hint.

oob_fill: CuTensorMapFloatOobFill

Out-of-bounds fill mode for float elements.

Trait Implementations

impl Clone for TmaEncodeTiledParams

fn clone(&self) -> TmaEncodeTiledParams

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TmaEncodeTiledParams

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

Formats the value using the given formatter.