add vpdpbusd(s), vpdpwssd(s)

This commit is contained in:
MITSUNARI Shigeo 2018-01-05 13:58:05 +09:00
parent 9c7451091e
commit 48499eb100
3 changed files with 57 additions and 4 deletions

View file

@ -408,9 +408,33 @@ CYBOZU_TEST_AUTO(vpdpbus)
vpdpbusd(ymm5|k3|T_z, ymm20, ptr [rax + 0x40]);
vpdpbusd(zmm5|k3|T_z, zmm20, ptr [rax + 0x40]);
vpdpbusd(xmm5|k3|T_z, xmm20, ptr_b [rax + 0x40]);
vpdpbusd(ymm5|k3|T_z, ymm20, ptr_b [rax + 0x40]);
vpdpbusd(zmm5|k3|T_z, zmm20, ptr_b [rax + 0x40]);
vpdpbusds(xmm5|k3|T_z, xmm20, ptr [rax + 0x40]);
vpdpbusds(ymm5|k3|T_z, ymm20, ptr [rax + 0x40]);
vpdpbusds(zmm5|k3|T_z, zmm20, ptr [rax + 0x40]);
vpdpbusds(xmm5|k3|T_z, xmm20, ptr_b [rax + 0x40]);
vpdpbusds(ymm5|k3|T_z, ymm20, ptr_b [rax + 0x40]);
vpdpbusds(zmm5|k3|T_z, zmm20, ptr_b [rax + 0x40]);
vpdpwssd(xmm5|k3|T_z, xmm20, ptr [rax + 0x40]);
vpdpwssd(ymm5|k3|T_z, ymm20, ptr [rax + 0x40]);
vpdpwssd(zmm5|k3|T_z, zmm20, ptr [rax + 0x40]);
vpdpwssd(xmm5|k3|T_z, xmm20, ptr_b [rax + 0x40]);
vpdpwssd(ymm5|k3|T_z, ymm20, ptr_b [rax + 0x40]);
vpdpwssd(zmm5|k3|T_z, zmm20, ptr_b [rax + 0x40]);
vpdpwssds(xmm5|k3|T_z, xmm20, ptr [rax + 0x40]);
vpdpwssds(ymm5|k3|T_z, ymm20, ptr [rax + 0x40]);
vpdpwssds(zmm5|k3|T_z, zmm20, ptr [rax + 0x40]);
vpdpwssds(xmm5|k3|T_z, xmm20, ptr_b [rax + 0x40]);
vpdpwssds(ymm5|k3|T_z, ymm20, ptr_b [rax + 0x40]);
vpdpwssds(zmm5|k3|T_z, zmm20, ptr_b [rax + 0x40]);
}
} c;
const uint8_t tbl[] = {
@ -418,9 +442,33 @@ CYBOZU_TEST_AUTO(vpdpbus)
0x62, 0xf2, 0x5d, 0xa3, 0x50, 0x68, 0x02,
0x62, 0xf2, 0x5d, 0xc3, 0x50, 0x68, 0x01,
0x62, 0xf2, 0x5d, 0x93, 0x50, 0x68, 0x10,
0x62, 0xf2, 0x5d, 0xb3, 0x50, 0x68, 0x10,
0x62, 0xf2, 0x5d, 0xd3, 0x50, 0x68, 0x10,
0x62, 0xf2, 0x5d, 0x83, 0x51, 0x68, 0x04,
0x62, 0xf2, 0x5d, 0xa3, 0x51, 0x68, 0x02,
0x62, 0xf2, 0x5d, 0xc3, 0x51, 0x68, 0x01,
0x62, 0xf2, 0x5d, 0x93, 0x51, 0x68, 0x10,
0x62, 0xf2, 0x5d, 0xb3, 0x51, 0x68, 0x10,
0x62, 0xf2, 0x5d, 0xd3, 0x51, 0x68, 0x10,
0x62, 0xf2, 0x5d, 0x83, 0x52, 0x68, 0x04,
0x62, 0xf2, 0x5d, 0xa3, 0x52, 0x68, 0x02,
0x62, 0xf2, 0x5d, 0xc3, 0x52, 0x68, 0x01,
0x62, 0xf2, 0x5d, 0x93, 0x52, 0x68, 0x10,
0x62, 0xf2, 0x5d, 0xb3, 0x52, 0x68, 0x10,
0x62, 0xf2, 0x5d, 0xd3, 0x52, 0x68, 0x10,
0x62, 0xf2, 0x5d, 0x83, 0x53, 0x68, 0x04,
0x62, 0xf2, 0x5d, 0xa3, 0x53, 0x68, 0x02,
0x62, 0xf2, 0x5d, 0xc3, 0x53, 0x68, 0x01,
0x62, 0xf2, 0x5d, 0x93, 0x53, 0x68, 0x10,
0x62, 0xf2, 0x5d, 0xb3, 0x53, 0x68, 0x10,
0x62, 0xf2, 0x5d, 0xd3, 0x53, 0x68, 0x10,
};
const size_t n = sizeof(tbl) / sizeof(tbl[0]);
CYBOZU_TEST_EQUAL(c.getSize(), n);