tileon.language

Memory Operations

load

Load data from memory with optional masking.

x = tl.load(ptr + offsets, mask=mask, other=0.0)

store

Store data to memory with optional masking.

tl.store(ptr + offsets, data, mask=mask)

make_block_ptr

Create a block pointer for efficient memory access.

tensor_descriptor

Create a tensor descriptor for complex memory layouts.

Indexing

arange

Create a range of indices.

offsets = tl.arange(0, BLOCK_SIZE)

program_id

Get the ID of the current program instance.

pid = tl.program_id(axis=0)

range / static_range

Create a static range for iteration.

full / zeros / zeros_like

Create filled tensors.

data = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)

Tensor Operations

broadcast / broadcast_to

Broadcast tensor to a new shape.

expand_dims / reshape / slice / split / squeeze / transpose / unsqueeze / view

Tensor shape manipulation operations.

join / cat

Concatenate tensors.

Math Operations

dot

Matrix multiplication for tiles.

c = tl.dot(a, b)

fma

Fused multiply-add.

add / sub / mul / div

Element-wise arithmetic operations.

Comparison

maximum / minimum

Element-wise max/min operations.

where

Conditional selection.

result = tl.where(condition, x, y)

Reduction

sum / max / min

Reduction operations across an axis.

result = tl.sum(x, axis=0)

argmax / argmin

Return indices of max/min values.

reduce

General reduction operation.

Sorting

sort / topk

Sorting operations.

Utility

cdiv

Ceiling division.

constexpr

Compile-time constant.

num_programs

Get total number of program instances.