Function tensor_expect_layout

ds4.c:3136–3175

            exp = 1;
            while ((mant & 0x0400) == 0) {
                mant <<= 1;
                exp--;
            }
            mant &= 0x03ff;
            bits = sign | ((exp + 127 - 15) << 23) | (mant << 13);
        }
    } else if (exp == 31) {
        bits = sign | 0x7f800000u | (mant << 13);
    } else {
        bits = sign | ((exp + 127 - 15) << 23) | (mant << 13);
    }

    float f;
    memcpy(&f, &bits, sizeof(f));
    return f;
#endif
}

static inline uint16_t f32_to_f16(float f) {
#if defined(__ARM_NEON)
    const float32x4_t fv = vdupq_n_f32(f);
    const float16x4_t hv = vcvt_f16_f32(fv);
    return vget_lane_u16(vreinterpret_u16_f16(hv), 0);
#else
    uint32_t bits;
    memcpy(&bits, &f, sizeof(bits));

    const uint32_t sign = (bits >> 16) & 0x8000u;
    int32_t exp = (int32_t)((bits >> 23) & 0xffu) - 127 + 15;
    uint32_t mant = bits & 0x7fffffu;

    if (exp <= 0) {
        if (exp < -10) return (uint16_t)sign;
        mant |= 0x800000u;
        const uint32_t shift = (uint32_t)(14 - exp);
        uint32_t half_mant = mant >> shift;
        const uint32_t round_bit = (mant >> (shift - 1)) & 1u;
        const uint32_t sticky = mant & ((1u << (shift - 1)) - 1u);
        if (round_bit && (sticky || (half_mant & 1u))) half_mant++;
        return (uint16_t)(sign | half_mant);
    }

    if (exp >= 31) {