Why use it

GFNI makes multiplication in GF(2⁸) and binary transforms of each byte direct vector operations. That is useful for finite-field kernels, bit reversal and changes of bit representation. An affine transform can express a whole byte's linear bit mapping without a sequence of shifts and masks.

Operations

import native.x86.gfni; provides these operations through native::native. native.x86 and native re-export them. Byte operands and results are simd<std::uint8_t,N,Arch> with N = 16, 32, 64. Matrices are simd<std::uint64_t,N/8,Arch>.

Operation	Result for each byte
gf2p8mulb<Arch>(a, b)	Field product
gf2p8affineqb<Arch, Imm8>(a, matrix)	Binary matrix product XOR Imm8
gf2p8affineinvqb<Arch, Imm8>(a, matrix)	Field inverse, then matrix product XOR Imm8

The field polynomial is x⁸ + x⁴ + x³ + x + 1 (0x11b). Inversion maps zero to zero. Each matrix qword supplies an independent 8×8 matrix for eight bytes. Numbering matrix bytes from the low end, result bit i is the parity of a_byte & matrix_byte[7-i], XOR bit i of the compile-time byte Imm8. 0x0102040810204080 is the identity matrix; 0x8040201008040201 reverses bits.

Each operation also has NAME_mask(src, k, a, b) and NAME_maskz(k, a, b) forms, with matrix replacing b for affine operations and the same template immediate. Masks are predicate<N,Arch>; inactive bytes retain src or clear.

Caveats

affineinv inverts each field element before the transform; it does not invert the matrix. Matrix byte order matters. Imm8 is in 0–255 and is shared by all result bytes.

Form	Runtime features	Caller target
Unmasked 128-bit	GFNI	"gfni"
Unmasked 256-bit	GFNI, AVX	"avx,gfni"
Unmasked 512-bit	GFNI, AVX512F	"avx512f,gfni"
Masked 128/256-bit	GFNI, AVX512F, AVX512BW, AVX512VL	"avx512bw,avx512vl,gfni"
Masked 512-bit	GFNI, AVX512F, AVX512BW	"avx512bw,gfni"

The unmasked 256-bit wrapper needs no AVX2. Masked wrappers require BW for Clang's byte-mask operations. GFNI is independent of the general AVX2 and AVX-512 profiles. Use target_features<native::x86>(target) and admit the whole matching caller target before entry. Legacy 128-bit execution needs XMM state; VEX also needs YMM state, and EVEX needs opmask and ZMM state.

Feature-bearing overloads are constexpr with native runtime paths. Weaker tags have consteval overloads when the chosen SSE2, AVX or AVX512F register storage exists. Results preserve their architecture tag; there is no runtime software fallback.

See Intel's GFNI instruction reference and Clang's GFNI header.