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

strided_load

def strided_load[dtype: DType, //, simd_width: SIMDLength, stride: Int, *, alignment: Int = Int((get_alignof Scalar[dtype], _current_target())), invariant: Bool = True](addr: Pointer[Scalar[dtype], address_space=addr.address_space], mask: SIMD[DType.bool, simd_width] = SIMD(fill=True)) -> SIMD[dtype, simd_width]

Loads simd_width values from addr with a compile-time stride.

Knowing the stride at compile time lets the two common cases avoid the gather used by the runtime-stride overload (which lowers to scattered per-element loads on NVIDIA GPUs): a contiguous stride == 1 load becomes a single coalesced vector load, and a stride == 0 broadcast becomes one scalar load splatted across lanes. Any other stride (or a partially-masked load) delegates to the runtime-stride overload.

Body authors use this for side-input loads (gamma/beta/cos/sin) along the reduced axis: inner axis gives stride=1, a non-inner axis gives stride=0 (one value per row, splatted across the tile).

invariant defaults True: addr is always mut=False here, and every current call site loads a read-only weight/side input (never aliased by an output of the same kernel), so the load-invariant hint is safe by construction. Pass invariant=False explicitly if a future call site's memory isn't provably invariant for the kernel's duration.

Parameters:

• ​dtype (DType): DType of the loaded value.
• ​simd_width (SIMDLength): The width of the SIMD vector.
• ​stride (Int): The compile-time stride, in elements, between lanes.
• ​alignment (Int): Alignment in bytes of the contiguous stride == 1 load; defaults to the element alignment. Pass the SIMD-natural alignment when addr is known to be vector-aligned to fold the load into one access.
• ​invariant (Bool): Whether the memory is load invariant.

Args:

• ​addr (Pointer[Scalar[dtype], address_space=addr.address_space]): The memory location to load data from.
• ​mask (SIMD[DType.bool, simd_width]): A binary vector which prevents memory access to certain lanes of the result.

Returns:

SIMD[dtype, simd_width]: A vector containing the loaded data.