vdpphps(xm1, xm2, xm3);
vdpphps(xm1, xm2, ptr[rax+128]);
vdpphps(xm1, xm2, ptr_b[rax+128]);

vdpphps(ym1, ym2, ym3);
vdpphps(ym1, ym2, ptr[rax+128]);
vdpphps(ym1, ym2, ptr_b[rax+128]);

vdpphps(zm1, zm2, zm3);
vdpphps(zm1, zm2, ptr[rax+128]);
vdpphps(zm1, zm2, ptr_b[rax+128]);

vmpsadbw(xm1, xm3, xm15, 3);
vmpsadbw(xm1|T_z, xm4, ptr[rax+128], 5);

vmpsadbw(ym1|k4, ym3, ym15, 3);
vmpsadbw(ym1, ym4, ptr[rax+128], 5);

vmpsadbw(zm1|k4, zm3, zm15, 3);
vmpsadbw(zm1, zm4, ptr[rax+128], 5);