FORCE_INLINE __m128i _mm_mpsadbw_epu8()

in common/checksum/sse2neon.h [7435:7514]

• 72 lines of code
• 10 McCabe index (conditional complexity)

FORCE_INLINE __m128i _mm_mpsadbw_epu8(__m128i a, __m128i b, const int imm)
{
    uint8x16_t _a, _b;

    switch (imm & 0x4) {
    case 0:
        // do nothing
        _a = vreinterpretq_u8_m128i(a);
        break;
    case 4:
        _a = vreinterpretq_u8_u32(vextq_u32(vreinterpretq_u32_m128i(a),
                                            vreinterpretq_u32_m128i(a), 1));
        break;
    default:
#if defined(__GNUC__) || defined(__clang__)
        __builtin_unreachable();
#elif defined(_MSC_VER)
        __assume(0);
#endif
        break;
    }

    switch (imm & 0x3) {
    case 0:
        _b = vreinterpretq_u8_u32(
            vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 0)));
        break;
    case 1:
        _b = vreinterpretq_u8_u32(
            vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 1)));
        break;
    case 2:
        _b = vreinterpretq_u8_u32(
            vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 2)));
        break;
    case 3:
        _b = vreinterpretq_u8_u32(
            vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 3)));
        break;
    default:
#if defined(__GNUC__) || defined(__clang__)
        __builtin_unreachable();
#elif defined(_MSC_VER)
        __assume(0);
#endif
        break;
    }

    int16x8_t c04, c15, c26, c37;
    uint8x8_t low_b = vget_low_u8(_b);
    c04 = vreinterpretq_s16_u16(vabdl_u8(vget_low_u8(_a), low_b));
    uint8x16_t _a_1 = vextq_u8(_a, _a, 1);
    c15 = vreinterpretq_s16_u16(vabdl_u8(vget_low_u8(_a_1), low_b));
    uint8x16_t _a_2 = vextq_u8(_a, _a, 2);
    c26 = vreinterpretq_s16_u16(vabdl_u8(vget_low_u8(_a_2), low_b));
    uint8x16_t _a_3 = vextq_u8(_a, _a, 3);
    c37 = vreinterpretq_s16_u16(vabdl_u8(vget_low_u8(_a_3), low_b));
#if defined(__aarch64__) || defined(_M_ARM64)
    // |0|4|2|6|
    c04 = vpaddq_s16(c04, c26);
    // |1|5|3|7|
    c15 = vpaddq_s16(c15, c37);

    int32x4_t trn1_c =
        vtrn1q_s32(vreinterpretq_s32_s16(c04), vreinterpretq_s32_s16(c15));
    int32x4_t trn2_c =
        vtrn2q_s32(vreinterpretq_s32_s16(c04), vreinterpretq_s32_s16(c15));
    return vreinterpretq_m128i_s16(vpaddq_s16(vreinterpretq_s16_s32(trn1_c),
                                              vreinterpretq_s16_s32(trn2_c)));
#else
    int16x4_t c01, c23, c45, c67;
    c01 = vpadd_s16(vget_low_s16(c04), vget_low_s16(c15));
    c23 = vpadd_s16(vget_low_s16(c26), vget_low_s16(c37));
    c45 = vpadd_s16(vget_high_s16(c04), vget_high_s16(c15));
    c67 = vpadd_s16(vget_high_s16(c26), vget_high_s16(c37));

    return vreinterpretq_m128i_s16(
        vcombine_s16(vpadd_s16(c01, c23), vpadd_s16(c45, c67)));
#endif
}