xbyak/test/avx10/minmax.txt

67 lines
2.2 KiB
Text
Raw Normal View History

2024-10-13 12:48:25 -07:00
vminmaxnepbf16(xm1|k3|T_z, xm2, xm3, 5);
vminmaxnepbf16(xm1|k3|T_z, xm2, ptr[rax+128], 5);
vminmaxnepbf16(xm1|k3|T_z, xm2, ptr_b[rax+128], 5);
vminmaxnepbf16(ym1|k3|T_z, ym2, ym3, 5);
vminmaxnepbf16(ym1|k3|T_z, ym2, ptr[rax+128], 5);
vminmaxnepbf16(ym1|k3|T_z, ym2, ptr_b[rax+128], 5);
vminmaxnepbf16(zm1|k3|T_z, zm2, zm3, 5);
vminmaxnepbf16(zm1|k3|T_z, zm2, ptr[rax+128], 5);
vminmaxnepbf16(zm1|k3|T_z, zm2, ptr_b[rax+128], 5);
//
vminmaxpd(xm1|k3|T_z, xm2, xm3, 5);
vminmaxpd(xm1|k3|T_z, xm2, ptr[rax+128], 5);
vminmaxpd(xm1|k3|T_z, xm2, ptr_b[rax+128], 5);
vminmaxpd(ym1|k3|T_z, ym2, ym3, 5);
vminmaxpd(ym1|k3|T_z, ym2, ym3|T_sae, 5);
vminmaxpd(ym1|k3|T_z, ym2, ptr[rax+128], 5);
vminmaxpd(ym1|k3|T_z, ym2, ptr_b[rax+128], 5);
vminmaxpd(zm1|k3|T_z, zm2, zm3, 5);
vminmaxpd(zm1|k3|T_z, zm2, zm3|T_sae, 5);
vminmaxpd(zm1|k3|T_z, zm2, ptr[rax+128], 5);
vminmaxpd(zm1|k3|T_z, zm2, ptr_b[rax+128], 5);
//
vminmaxph(xm1|k3|T_z, xm2, xm3, 5);
vminmaxph(xm1|k3|T_z, xm2, ptr[rax+128], 5);
vminmaxph(xm1|k3|T_z, xm2, ptr[rax+128], 5);
vminmaxph(xm1|k3|T_z, xm2, ptr_b[rax+128], 5);
vminmaxph(ym1|k3|T_z, ym2, ym3, 5);
vminmaxph(ym1|k3|T_z, ym2, ym3|T_sae, 5);
vminmaxph(ym1|k3|T_z, ym2, ptr[rax+128], 5);
vminmaxph(ym1|k3|T_z, ym2, ptr_b[rax+128], 5);
vminmaxph(zm1|k3|T_z, zm2, zm3, 5);
vminmaxph(zm1|k3|T_z, zm2, zm3|T_sae, 5);
vminmaxph(zm1|k3|T_z, zm2, ptr[rax+128], 5);
vminmaxph(zm1|k3|T_z, zm2, ptr_b[rax+128], 5);
//
vminmaxps(xm1|k3|T_z, xm2, xm3, 5);
vminmaxps(xm1|k3|T_z, xm2, ptr[rax+128], 5);
vminmaxps(xm1|k3|T_z, xm2, ptr_b[rax+128], 5);
vminmaxps(ym1|k3|T_z, ym2, ym3, 5);
vminmaxps(ym1|k3|T_z, ym2, ym3|T_sae, 5);
vminmaxps(ym1|k3|T_z, ym2, ptr[rax+128], 5);
vminmaxps(ym1|k3|T_z, ym2, ptr_b[rax+128], 5);
vminmaxps(zm1|k3|T_z, zm2, zm3, 5);
vminmaxps(zm1|k3|T_z, zm2, zm3|T_sae, 5);
vminmaxps(zm1|k3|T_z, zm2, ptr[rax+128], 5);
vminmaxps(zm1|k3|T_z, zm2, ptr_b[rax+128], 5);
//
vminmaxsd(xm1|k3|T_z, xm2, xm3, 5);
vminmaxsd(xm1|k3|T_z, xm2, xm3|T_sae, 5);
vminmaxsd(xm1|k3|T_z, xm2, ptr[rax+128], 5);
//
vminmaxsh(xm1|k3|T_z, xm2, xm3, 5);
vminmaxsh(xm1|k3|T_z, xm2, xm3|T_sae, 5);
vminmaxsh(xm1|k3|T_z, xm2, ptr[rax+128], 5);
//
vminmaxss(xm1|k3|T_z, xm2, xm3, 5);
vminmaxss(xm1|k3|T_z, xm2, xm3|T_sae, 5);
vminmaxss(xm1|k3|T_z, xm2, ptr[rax+128], 5);