For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

tile_alignment

def tile_alignment[dtype: DType, ws: Int, target: StringSpan[ImmStaticOrigin]]() -> Int

Alignment, in elements, a row-wise body's store — or its input_fn's load — may assume for a width-ws tile of dtype on target.

The unit is elements, not bytes: every caller forwards this value into the element_alignment parameter of a tensor load/store, which multiplies it by align_of[dtype]() to get the byte alignment it promises the backend. Returning a byte count here therefore squares the intended alignment, and an alignment the address does not keep is a fault rather than a missed optimization — a float64 tile promised 64 bytes loads through vmovapd.

Single source of truth for the rule. 1 on CPU (element-natural: the CPU tiers walk a row at the tensor's own stride, and an input tensor may be an imported buffer, so neither a tile base nor the step between tiles is guaranteed SIMD-aligned). ws on GPU (SIMD-natural — ws * align_of[dtype]() bytes: folds the access into one vector transaction (LDG.128/STG.128), safe because a body only ever gets a ws > 1 tile on a SIMD-aligned offset (block tier requires row_size % simd == 0; the warp tier matches via the same guard, else falls through to the block tier)). The two coincide at ws == 1.

Both Context.element_alignment (the store) and the public-op wrappers' synthesized input_fn (the load) call this, so the rule lives in one place.

Parameters:

• ​dtype (DType): The tile's element dtype.
• ​ws (Int): The tile's SIMD width.
• ​target (StringSpan[ImmStaticOrigin]): "cpu" or "gpu".

Returns:

Int: The alignment in elements.