Chunked Processing

The standard pattern for processing large arrays with SIMD: iterate in fixed-size chunks, handle the remainder with scalar code.

See the complete #[magetypes] image-plane loop and reusable generic helper. Dispatch once, iterate array chunks, and handle the remainder explicitly. For a reduction, accumulate in a vector inside that same context and call reduce_add() once after the loop; floating-point association can differ by ISA.

Alignment Tips

Align your structs

Only if measurements require stronger alignment, an owned type can request it:

#[repr(C, align(32))]
struct AlignedData {
    values: [f32; 8],
}

Allocate aligned memory safely

When a measured consumer needs 32-byte alignment, let Rust allocate an aligned type and own its lifetime:

#[repr(C, align(32))]
struct AlignedBlock([f32; 8]);
let blocks = vec![AlignedBlock([0.0; 8]), AlignedBlock([0.0; 8])];
assert_eq!(blocks.as_ptr() as usize % 32, 0);

The vector drops its allocation normally, and every element is initialized. Ordinary magetypes loads do not require this stronger alignment.

Performance tips

Keep the loop inside one generated context; keep intermediate vectors in registers; retain a scalar tail. Test cache-resident and streaming working sets. Use measurements to choose layout and width rather than assuming that alignment, gather avoidance, or non-temporal stores always improve performance.