diff options
| author | MITSUNARI Shigeo <[email protected]> | 2024-10-13 15:27:05 +0900 |
|---|---|---|
| committer | MITSUNARI Shigeo <[email protected]> | 2024-10-13 15:27:05 +0900 |
| commit | f6c66cf6b81f7a063a930cdfc0a62c68e6e2d0fc (patch) | |
| tree | 1f0890d9bf310f34d5f0ef7132d13ea613187480 /test | |
| parent | f3f2dd2d748859fd4438ab596950ba52769607a4 (diff) | |
| download | xbyak-f6c66cf6b81f7a063a930cdfc0a62c68e6e2d0fc.tar.gz xbyak-f6c66cf6b81f7a063a930cdfc0a62c68e6e2d0fc.zip | |
vpdpb[su,uu,ss]d[,s] support avx10.2
Diffstat (limited to 'test')
| -rw-r--r-- | test/avx10/misc.txt | 65 |
1 files changed, 63 insertions, 2 deletions
diff --git a/test/avx10/misc.txt b/test/avx10/misc.txt index 8993107..380e9a9 100644 --- a/test/avx10/misc.txt +++ b/test/avx10/misc.txt @@ -9,7 +9,7 @@ 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); @@ -18,7 +18,7 @@ vmpsadbw(ym1, ym4, ptr[rax+128], 5); vmpsadbw(zm1|k4, zm3, zm15, 3); vmpsadbw(zm1, zm4, ptr[rax+128], 5); - +// vpdpbssd(xm1, xm2, xm3); vpdpbssd(xm1, xm2, ptr[rax+128]); vpdpbssd(xm1, xm2, ptr_b[rax+128]); @@ -30,3 +30,64 @@ vpdpbssd(ym1, ym2, ptr_b[rax+128]); vpdpbssd(zm1, zm2, zm3); vpdpbssd(zm1, zm2, ptr[rax+128]); vpdpbssd(zm1, zm2, ptr_b[rax+128]); +// +vpdpbssds(xm1, xm2, xm3); +vpdpbssds(xm1, xm2, ptr[rax+128]); +vpdpbssds(xm1, xm2, ptr_b[rax+128]); + +vpdpbssds(ym1, ym2, ym3); +vpdpbssds(ym1, ym2, ptr[rax+128]); +vpdpbssds(ym1, ym2, ptr_b[rax+128]); + +vpdpbssds(zm1, zm2, zm3); +vpdpbssds(zm1, zm2, ptr[rax+128]); +vpdpbssds(zm1, zm2, ptr_b[rax+128]); +// +vpdpbsud(xm1, xm2, xm3); +vpdpbsud(xm1, xm2, ptr[rax+128]); +vpdpbsud(xm1, xm2, ptr_b[rax+128]); + +vpdpbsud(ym1, ym2, ym3); +vpdpbsud(ym1, ym2, ptr[rax+128]); +vpdpbsud(ym1, ym2, ptr_b[rax+128]); + +vpdpbsud(zm1, zm2, zm3); +vpdpbsud(zm1, zm2, ptr[rax+128]); +vpdpbsud(zm1, zm2, ptr_b[rax+128]); +// +vpdpbsuds(xm1, xm2, xm3); +vpdpbsuds(xm1, xm2, ptr[rax+128]); +vpdpbsuds(xm1, xm2, ptr_b[rax+128]); + +vpdpbsuds(ym1, ym2, ym3); +vpdpbsuds(ym1, ym2, ptr[rax+128]); +vpdpbsuds(ym1, ym2, ptr_b[rax+128]); + +vpdpbsuds(zm1, zm2, zm3); +vpdpbsuds(zm1, zm2, ptr[rax+128]); +vpdpbsuds(zm1, zm2, ptr_b[rax+128]); + +// +vpdpbuud(xm1, xm2, xm3); +vpdpbuud(xm1, xm2, ptr[rax+128]); +vpdpbuud(xm1, xm2, ptr_b[rax+128]); + +vpdpbuud(ym1, ym2, ym3); +vpdpbuud(ym1, ym2, ptr[rax+128]); +vpdpbuud(ym1, ym2, ptr_b[rax+128]); + +vpdpbuud(zm1, zm2, zm3); +vpdpbuud(zm1, zm2, ptr[rax+128]); +vpdpbuud(zm1, zm2, ptr_b[rax+128]); +// +vpdpbuuds(xm1, xm2, xm3); +vpdpbuuds(xm1, xm2, ptr[rax+128]); +vpdpbuuds(xm1, xm2, ptr_b[rax+128]); + +vpdpbuuds(ym1, ym2, ym3); +vpdpbuuds(ym1, ym2, ptr[rax+128]); +vpdpbuuds(ym1, ym2, ptr_b[rax+128]); + +vpdpbuuds(zm1, zm2, zm3); +vpdpbuuds(zm1, zm2, ptr[rax+128]); +vpdpbuuds(zm1, zm2, ptr_b[rax+128]); |