vaddnepbf16(xm1, xm2, xm3);
vaddnepbf16(ym1|k1, ym2, ptr[rax+128]);
vaddnepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vaddnepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vdivnepbf16(xm1, xm2, xm3);
vdivnepbf16(ym1|k1, ym2, ptr[rax+128]);
vdivnepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vdivnepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vmaxpbf16(xm1, xm2, xm3);
vmaxpbf16(ym1|k1, ym2, ptr[rax+128]);
vmaxpbf16(ym1|k1, ym2, ptr_b[rax+128]);
vmaxpbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vminpbf16(xm1, xm2, xm3);
vminpbf16(ym1|k1, ym2, ptr[rax+128]);
vminpbf16(ym1|k1, ym2, ptr_b[rax+128]);
vminpbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vmulnepbf16(xm1, xm2, xm3);
vmulnepbf16(ym1|k1, ym2, ptr[rax+128]);
vmulnepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vmulnepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vscalefpbf16(xm1, xm2, xm3);
vscalefpbf16(ym1|k1, ym2, ptr[rax+128]);
vscalefpbf16(ym1|k1, ym2, ptr_b[rax+128]);
vscalefpbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vsubnepbf16(xm1, xm2, xm3);
vsubnepbf16(ym1|k1, ym2, ptr[rax+128]);
vsubnepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vsubnepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);
// madd
vfmadd132nepbf16(xm1, xm2, xm3);
vfmadd132nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfmadd132nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfmadd132nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vfmadd213nepbf16(xm1, xm2, xm3);
vfmadd213nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfmadd213nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfmadd213nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vfmadd231nepbf16(xm1, xm2, xm3);
vfmadd231nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfmadd231nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfmadd231nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);
// nmadd
vfnmadd132nepbf16(xm1, xm2, xm3);
vfnmadd132nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfnmadd132nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfnmadd132nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vfnmadd213nepbf16(xm1, xm2, xm3);
vfnmadd213nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfnmadd213nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfnmadd213nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vfnmadd231nepbf16(xm1, xm2, xm3);
vfnmadd231nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfnmadd231nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfnmadd231nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);
// msub
vfmsub132nepbf16(xm1, xm2, xm3);
vfmsub132nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfmsub132nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfmsub132nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vfmsub213nepbf16(xm1, xm2, xm3);
vfmsub213nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfmsub213nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfmsub213nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vfmsub231nepbf16(xm1, xm2, xm3);
vfmsub231nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfmsub231nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfmsub231nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);
// nmsub
vfnmsub132nepbf16(xm1, xm2, xm3);
vfnmsub132nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfnmsub132nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfnmsub132nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vfnmsub213nepbf16(xm1, xm2, xm3);
vfnmsub213nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfnmsub213nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfnmsub213nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vfnmsub231nepbf16(xm1, xm2, xm3);
vfnmsub231nepbf16(ym1|k1, ym2, ptr[rax+128]);
vfnmsub231nepbf16(ym1|k1, ym2, ptr_b[rax+128]);
vfnmsub231nepbf16(zm1|k2|T_z, zm2, ptr_b[rax+128]);

vcmppbf16(k1, xm5, xm4, 5);
vcmppbf16(k2, ym5, ym4, 6);
vcmppbf16(k3, ym15, ptr_b[rax+128], 7);
vcmppbf16(k4, zm30, zm20, 8);
vcmppbf16(k5, zm1, ptr[rax+128], 9);
vcmppbf16(k6, zm10, ptr_b[rax+128], 10);

vfpclasspbf16(k1, xm4, 5);
vfpclasspbf16(k2|k5, ym4, 6);
vfpclasspbf16(k3|k5, zm20, 7);
vfpclasspbf16(k3|k5, xword[rax+128], 8);
vfpclasspbf16(k3, xword_b[rax+128], 9);
vfpclasspbf16(k5|k5, yword[rax+128], 10);
vfpclasspbf16(k6|k5, yword_b[rax+128], 11);
vfpclasspbf16(k7|k5, zword[rax+128], 12);
vfpclasspbf16(k7|k5, zword_b[rax+128], 13);

vcomsbf16(xm2, xm3);
vcomsbf16(xm2, ptr[rax+128]);

//vgetexppbf16(xm1|k3, xmm2);
//vgetexppbf16(xm1|k3, ptr[rax+128]);
//vgetexppbf16(xm1|k3, ptr_b[rax+128]);

//vgetexppbf16(ym1|k3, ymm2);
//vgetexppbf16(ym1|k3, ptr[rax+128]);
//vgetexppbf16(ym1|k3, ptr_b[rax+128]);

//vgetexppbf16(zm1|k3, zmm2);
//vgetexppbf16(zm1|k3, ptr[rax+128]);
//vgetexppbf16(zm1|k3, ptr_b[rax+128]);

vgetmantpbf16(xm1|k3, xmm2, 3);
vgetmantpbf16(xm1|k3, ptr[rax+128], 5);
vgetmantpbf16(xm1|k3, ptr_b[rax+128], 9);

vgetmantpbf16(ym1|k3, ymm2, 3);
vgetmantpbf16(ym1|k3, ptr[rax+128], 5);
vgetmantpbf16(ym1|k3, ptr_b[rax+128], 9);

vgetmantpbf16(zm1|k3, zmm2, 3);
vgetmantpbf16(zm1|k3, ptr[rax+128], 5);
vgetmantpbf16(zm1|k3, ptr_b[rax+128], 9);

vrcppbf16(xm1|k5, xm2);
vrcppbf16(xm1|k5, ptr[rcx+128]);
vrcppbf16(xm1|k5, ptr_b[rcx+128]);

vrcppbf16(ym1|k5, ym2);
vrcppbf16(ym1|k5, ptr[rcx+128]);
vrcppbf16(ym1|k5, ptr_b[rcx+128]);

vrcppbf16(zm1|k5, zm2);
vrcppbf16(zm1|k5, ptr[rcx+128]);
vrcppbf16(zm1|k5, ptr_b[rcx+128]);

vreducenepbf16(xm1|k4, xm2, 1);
vreducenepbf16(xm1|k4, ptr[rax+128], 1);
vreducenepbf16(xm1|k4, ptr_b[rax+128], 1);

vreducenepbf16(ym1|k4, ym2, 1);
vreducenepbf16(ym1|k4, ptr[rax+128], 1);
vreducenepbf16(ym1|k4, ptr_b[rax+128], 1);

vreducenepbf16(zm1|k4, zm2, 1);
vreducenepbf16(zm1|k4, ptr[rax+128], 1);
vreducenepbf16(zm1|k4, ptr_b[rax+128], 1);

vrndscalenepbf16(xm1|k4, xm2, 1);
vrndscalenepbf16(xm1|k4, ptr[rax+128], 1);
vrndscalenepbf16(xm1|k4, ptr_b[rax+128], 1);

vrndscalenepbf16(ym1|k4, ym2, 1);
vrndscalenepbf16(ym1|k4, ptr[rax+128], 1);
vrndscalenepbf16(ym1|k4, ptr_b[rax+128], 1);

vrndscalenepbf16(zm1|k4, zm2, 1);
vrndscalenepbf16(zm1|k4, ptr[rax+128], 1);
vrndscalenepbf16(zm1|k4, ptr_b[rax+128], 1);

vrsqrtpbf16(xm1|k5, xm2);
vrsqrtpbf16(xm1|k5, ptr[rcx+128]);
vrsqrtpbf16(xm1|k5, ptr_b[rcx+128]);

vrsqrtpbf16(ym1|k5, ym2);
vrsqrtpbf16(ym1|k5, ptr[rcx+128]);
vrsqrtpbf16(ym1|k5, ptr_b[rcx+128]);

vrsqrtpbf16(zm1|k5, zm2);
vrsqrtpbf16(zm1|k5, ptr[rcx+128]);
vrsqrtpbf16(zm1|k5, ptr_b[rcx+128]);

vscalefpbf16(xm1|k5, xm5, xm2);
vscalefpbf16(xm1|k5, xm5, ptr[rcx+128]);
vscalefpbf16(xm1|k5, xm5, ptr_b[rcx+128]);

vscalefpbf16(ym1|k5, ym9, ym2);
vscalefpbf16(ym1|k5, ym9, ptr[rcx+128]);
vscalefpbf16(ym1|k5, ym9, ptr_b[rcx+128]);

vscalefpbf16(zm1|k5, zm30, zm2);
vscalefpbf16(zm1|k5, zm30, ptr[rcx+128]);
vscalefpbf16(zm1|k5, zm30, ptr_b[rcx+128]);

vsqrtnepbf16(xm5|k3, xmm4);
vsqrtnepbf16(xm5|k3, ptr[rax+128]);
vsqrtnepbf16(xm5|k3, ptr_b[rax+128]);

vsqrtnepbf16(ym5|k3, ymm4);
vsqrtnepbf16(ym5|k3, ptr[rax+128]);
vsqrtnepbf16(ym5|k3, ptr_b[rax+128]);

vsqrtnepbf16(zm5|k3, zmm4);
vsqrtnepbf16(zm5|k3, ptr[rax+128]);
vsqrtnepbf16(zm5|k3, ptr_b[rax+128]);