cpu_features.hpp File Reference

Runtime SIMD capability detection for the SRFM acceleration layer. More...

#include <cstdint>

Go to the source code of this file.

Namespaces
namespace	srfm
namespace	srfm::simd
namespace	srfm::simd::detail

Enumerations
enum class	srfm::simd::SimdLevel : std::uint8_t { srfm::simd::SCALAR = 0 , srfm::simd::SSE42 = 1 , srfm::simd::AVX2 = 2 , srfm::simd::AVX512F = 3 }
	Ordered enumeration of SIMD capability tiers. More...

Functions
void	srfm::simd::detail::cpuid (std::uint32_t leaf, std::uint32_t subleaf, std::uint32_t out[4]) noexcept
std::uint64_t	srfm::simd::detail::xgetbv (std::uint32_t xcr) noexcept
	Read XCR0 (used to verify OS XSAVE support for YMM/ZMM state).
bool	srfm::simd::detail::os_saves_ymm () noexcept
	Probe OS support for YMM (AVX) register save/restore.
bool	srfm::simd::detail::os_saves_zmm () noexcept
	Probe OS support for ZMM (AVX-512) register save/restore.
SimdLevel	srfm::simd::detect_simd_level () noexcept
bool	srfm::simd::has_avx512f () noexcept
	Returns true when AVX-512F is available on this CPU/OS.
bool	srfm::simd::has_avx2 () noexcept
	Returns true when AVX2 is available on this CPU/OS.
bool	srfm::simd::has_sse42 () noexcept
	Returns true when SSE 4.2 is available on this CPU/OS.
const char *	srfm::simd::simd_level_name (SimdLevel level) noexcept
	Human-readable name for a SimdLevel value.

Detailed Description

Runtime SIMD capability detection for the SRFM acceleration layer.

Module: include/srfm/simd/ Owner: AGT-08 (AVX-512 SIMD Acceleration) — 2026-03-01

Responsibility

Probe the executing CPU at start-up for the widest SIMD instruction set available, and expose the result as a strongly-typed enum so that the dispatch layer can select the correct kernel at zero per-call overhead.

Supported levels (highest wins): SCALAR — always available; pure C++ reference path SSE42 — SSE 4.2 (128-bit, 2 doubles per register) AVX2 — AVX2 (256-bit, 4 doubles per register) AVX512F — AVX-512F (512-bit, 8 doubles per register)

Guarantees

• detect_simd_level() is thread-safe: result is computed once. • No dynamic allocation, no exceptions, noexcept throughout. • Portable: supports GCC/Clang (__builtin_cpu_supports) and MSVC (__cpuid / __cpuidex).

NOT Responsible For

• Verifying OS support (XSAVE/XGETBV) — assumed present if CPUID advertises the feature and the OS is modern (≥ Win10 / Linux ≥ 4.x). • AVX-512 sub-features beyond F (VL, DQ, BW) — only F is required.

Definition in file cpu_features.hpp.