helion.language.grid

helion.language.grid(begin_or_end, end_or_none=None, /, step=None)

Iterate over individual indices of the given iteration space.

The key difference from tile() is that grid gives you scalar integer indices (torch.SymInt), while tile gives you Tile objects that load a slice of elements. Use tile in most cases. Use grid when you need explicit control over the launch grid or when processing one element at a time.

Semantics are equivalent to:

for i in hl.tile(...):
    # i is a Tile object, accesses multiple elements
    data = tensor[i]  # loads slice of elements (1D tensor)

vs:

for i in hl.grid(...):
    # i is a scalar index, accesses single element
    data = tensor[i]  # loads single element (0D scalar)

When used at the top level of a function, this becomes the grid of the kernel. Otherwise, it becomes a loop in the output kernel.

Parameters:

• begin_or_end (int | Tensor | Sequence[int | Tensor]) – If 2+ positional args provided, the start of iteration space. Otherwise, the end of iteration space.

• end_or_none (int | Tensor | Sequence[int | Tensor] | None) – If 2+ positional args provided, the end of iteration space.

• step (object) – Step size for iteration (default: 1)

Returns:

Iterator over scalar indices

Return type:

Iterator[torch.SymInt] or Iterator[Sequence[torch.SymInt]]

See also

• tile(): For processing multiple elements at once

• tile_index(): For getting tile indices

• arange(): For creating index sequences

Note

Similar to range() with multiple forms:

• grid(end) iterates from 0 to end-1, step 1

• grid(begin, end) iterates from begin to end-1, step 1

• grid(begin, end, step) iterates from begin to end-1, given step

• grid(end, step=step) iterates from 0 to end-1, given step

Use tile in most cases. Use grid when you need explicit control over the launch grid.