in common/checksum/sse2neon.h [915:1005]
static uint64x2_t _sse2neon_vmull_p64(uint64x1_t _a, uint64x1_t _b)
{
poly8x8_t a = vreinterpret_p8_u64(_a);
poly8x8_t b = vreinterpret_p8_u64(_b);
// Masks
uint8x16_t k48_32 = vcombine_u8(vcreate_u8(0x0000ffffffffffff),
vcreate_u8(0x00000000ffffffff));
uint8x16_t k16_00 = vcombine_u8(vcreate_u8(0x000000000000ffff),
vcreate_u8(0x0000000000000000));
// Do the multiplies, rotating with vext to get all combinations
uint8x16_t d = vreinterpretq_u8_p16(vmull_p8(a, b)); // D = A0 * B0
uint8x16_t e =
vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 1))); // E = A0 * B1
uint8x16_t f =
vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 1), b)); // F = A1 * B0
uint8x16_t g =
vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 2))); // G = A0 * B2
uint8x16_t h =
vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 2), b)); // H = A2 * B0
uint8x16_t i =
vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 3))); // I = A0 * B3
uint8x16_t j =
vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 3), b)); // J = A3 * B0
uint8x16_t k =
vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 4))); // L = A0 * B4
// Add cross products
uint8x16_t l = veorq_u8(e, f); // L = E + F
uint8x16_t m = veorq_u8(g, h); // M = G + H
uint8x16_t n = veorq_u8(i, j); // N = I + J
// Interleave. Using vzip1 and vzip2 prevents Clang from emitting TBL
// instructions.
#if defined(__aarch64__)
uint8x16_t lm_p0 = vreinterpretq_u8_u64(
vzip1q_u64(vreinterpretq_u64_u8(l), vreinterpretq_u64_u8(m)));
uint8x16_t lm_p1 = vreinterpretq_u8_u64(
vzip2q_u64(vreinterpretq_u64_u8(l), vreinterpretq_u64_u8(m)));
uint8x16_t nk_p0 = vreinterpretq_u8_u64(
vzip1q_u64(vreinterpretq_u64_u8(n), vreinterpretq_u64_u8(k)));
uint8x16_t nk_p1 = vreinterpretq_u8_u64(
vzip2q_u64(vreinterpretq_u64_u8(n), vreinterpretq_u64_u8(k)));
#else
uint8x16_t lm_p0 = vcombine_u8(vget_low_u8(l), vget_low_u8(m));
uint8x16_t lm_p1 = vcombine_u8(vget_high_u8(l), vget_high_u8(m));
uint8x16_t nk_p0 = vcombine_u8(vget_low_u8(n), vget_low_u8(k));
uint8x16_t nk_p1 = vcombine_u8(vget_high_u8(n), vget_high_u8(k));
#endif
// t0 = (L) (P0 + P1) << 8
// t1 = (M) (P2 + P3) << 16
uint8x16_t t0t1_tmp = veorq_u8(lm_p0, lm_p1);
uint8x16_t t0t1_h = vandq_u8(lm_p1, k48_32);
uint8x16_t t0t1_l = veorq_u8(t0t1_tmp, t0t1_h);
// t2 = (N) (P4 + P5) << 24
// t3 = (K) (P6 + P7) << 32
uint8x16_t t2t3_tmp = veorq_u8(nk_p0, nk_p1);
uint8x16_t t2t3_h = vandq_u8(nk_p1, k16_00);
uint8x16_t t2t3_l = veorq_u8(t2t3_tmp, t2t3_h);
// De-interleave
#if defined(__aarch64__)
uint8x16_t t0 = vreinterpretq_u8_u64(
vuzp1q_u64(vreinterpretq_u64_u8(t0t1_l), vreinterpretq_u64_u8(t0t1_h)));
uint8x16_t t1 = vreinterpretq_u8_u64(
vuzp2q_u64(vreinterpretq_u64_u8(t0t1_l), vreinterpretq_u64_u8(t0t1_h)));
uint8x16_t t2 = vreinterpretq_u8_u64(
vuzp1q_u64(vreinterpretq_u64_u8(t2t3_l), vreinterpretq_u64_u8(t2t3_h)));
uint8x16_t t3 = vreinterpretq_u8_u64(
vuzp2q_u64(vreinterpretq_u64_u8(t2t3_l), vreinterpretq_u64_u8(t2t3_h)));
#else
uint8x16_t t1 = vcombine_u8(vget_high_u8(t0t1_l), vget_high_u8(t0t1_h));
uint8x16_t t0 = vcombine_u8(vget_low_u8(t0t1_l), vget_low_u8(t0t1_h));
uint8x16_t t3 = vcombine_u8(vget_high_u8(t2t3_l), vget_high_u8(t2t3_h));
uint8x16_t t2 = vcombine_u8(vget_low_u8(t2t3_l), vget_low_u8(t2t3_h));
#endif
// Shift the cross products
uint8x16_t t0_shift = vextq_u8(t0, t0, 15); // t0 << 8
uint8x16_t t1_shift = vextq_u8(t1, t1, 14); // t1 << 16
uint8x16_t t2_shift = vextq_u8(t2, t2, 13); // t2 << 24
uint8x16_t t3_shift = vextq_u8(t3, t3, 12); // t3 << 32
// Accumulate the products
uint8x16_t cross1 = veorq_u8(t0_shift, t1_shift);
uint8x16_t cross2 = veorq_u8(t2_shift, t3_shift);
uint8x16_t mix = veorq_u8(d, cross1);
uint8x16_t r = veorq_u8(mix, cross2);
return vreinterpretq_u64_u8(r);
}