sysl.vectors

reference/vectors.md § Rearranging lanes describes shuffle itself.

Each is a const function answering an array of lane numbers, which is the form shuffle folds while compiling: a.shuffle(b, zip_lo[W]()) is one shufflevector whose list is the instruction’s own operand, so a kernel generic in its width interleaves at four lanes and at sixteen from one body, and the list costs nothing at run time. They are lists rather than methods because a list is what may be named, combined and handed on; the method that applies one is the compiler’s.

W is always the width of the vectors being shuffled, never the width of the answer, and the second vector’s lanes are numbered W to 2W - 1 as shuffle numbers them. The lists that take a pair say so; the rest pick from one vector.

Index

reverse rotate unzip_even unzip_odd zip_hi zip_lo

Functions

reverse

const reverse[const W: usize]() -> [W]usize

One vector’s lanes back to front. rev64 and an ext on AArch64.

rotate

const rotate[const W: usize, const K: usize]() -> [W]usize

One vector’s lanes moved K toward lane zero, the ones that fall off coming round at the top: lane i of the answer is lane (i + K) % W. ext on AArch64.

unzip_even

const unzip_even[const W: usize]() -> [W]usize

Every even lane of a pair, in order: [a0, a2, …, b0, b2, …] – the inverse of the two zips. uzp1 on AArch64.

unzip_odd

const unzip_odd[const W: usize]() -> [W]usize

Every odd lane of a pair, in order: [a1, a3, …, b1, b3, …]. uzp2 on AArch64.

zip_hi

const zip_hi[const W: usize]() -> [W]usize

The high halves of two vectors, interleaved: [a(W/2), b(W/2), …], W lanes from a pair of W. zip2 on AArch64.

zip_lo

const zip_lo[const W: usize]() -> [W]usize

The low halves of two vectors, interleaved: [a0, b0, a1, b1, …], W lanes from a pair of W. zip1 on AArch64, unpcklps on x86.