9848     __m128i ld128;  variable
9849     ld128 = vld1q_u8(ptr); //merge two 64-bits in 128 bit
9850     ld128 =  _mm_shuffle_epi8(ld128, *(__m128i*)mask8_16_even_odd);
9851     vst1q_u8((v.val), ld128); //  v.val[1] = _mm_shuffle_epi32(v.val[0], _SWAP_HI_LOW32);
9859     __m128i ld128;  variable
9860     ld128 = vld1q_u16(ptr); //merge two 64-bits in 128 bit
9861     ld128 = _mm_shuffle_epi8(ld128, *(__m128i*) mask8_32_even_odd);
9862     vst1q_u16((v.val), ld128);
9870     __m128i ld128;  variable
9871     ld128 = vld1q_u32(ptr); //merge two 64-bits in 128 bit
9872     ld128 = _mm_shuffle_epi32(ld128,  0 | (2 << 2) | (1 << 4) | (3 << 6));
9873     vst1q_u32((v.val), ld128);