//ad_im_128 = _mm_shufflelo_epi16(after_mf_00_128[0],_MM_SHUFFLE(2,3,0,1));//permutes IQs for the low 64 bits as [I_a0 Q_a1 I_a2 Q_a3]_64bits to [Q_a1 I_a0 Q_a3 I_a2]_64bits
//ad_im_128 = _mm_shufflehi_epi16(ad_im_128,_MM_SHUFFLE(2,3,0,1));//permutes IQs for the high 64 bits as [I_a0 Q_a1 I_a2 Q_a3]_64bits to [Q_a1 I_a0 Q_a3 I_a2]_64bits
//bc_im_128 = _mm_shufflelo_epi16(after_mf_01_128[0],_MM_SHUFFLE(2,3,0,1));//permutes IQs for the low 64 bits as [I_b0 Q_b1 I_b2 Q_b3]_64bits to [Q_b1 I_b0 Q_b3 I_b2]_64bits