Explicit AVX-512 batch kernels for the bulk encode/decode/arithmetic paths. More...

#include <cstddef>
#include <cstdint>
#include "morton/morton.hpp"

Include dependency graph for simd.hpp:

This graph shows which files directly or indirectly include this file:

Detailed Description

Explicit AVX-512 batch kernels for the bulk encode/decode/arithmetic paths.

Unlike the auto-vectorised loops in batch.hpp, these are hand-written with AVX-512 intrinsics ("magic-bits" bit spreading, libmorton-style) and process eight 64-bit codes per iteration.

They are compiled with a per-function __attribute__((target("avx512f"))), so the translation unit does not need -mavx512f and the rest of the binary stays portable. batch.hpp dispatches to them at runtime only when detail::cpu_has_avx512f() is true, so a single binary uses AVX-512 on a capable CPU and the scalar/AVX2 path elsewhere. Codes that are not 64 bits wide keep the scalar path. Validate changes under Intel SDE (sde64 -skx) — most dev/CI hosts lack AVX-512.

SPDX-License-Identifier: MIT