nb_kernel102_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,986 行 · 第 1/5 页
S
1,986 行
movlhps %xmm3,%xmm1 movaps %xmm1,nb102_jxH1(%esp) movhlps %xmm7,%xmm3 movaps %xmm4,%xmm6 movaps %xmm3,nb102_jyH1(%esp) movlhps %xmm5,%xmm4 movaps %xmm4,nb102_jxH2(%esp) movhlps %xmm6,%xmm5 movaps %xmm5,nb102_jyH2(%esp) movss 8(%esi,%eax,4),%xmm0 movss 20(%esi,%eax,4),%xmm1 movss 32(%esi,%eax,4),%xmm2 movss 8(%esi,%ecx,4),%xmm3 movss 20(%esi,%ecx,4),%xmm4 movss 32(%esi,%ecx,4),%xmm5 movhps 4(%esi,%ebx,4),%xmm0 movhps 16(%esi,%ebx,4),%xmm1 movhps 28(%esi,%ebx,4),%xmm2 movhps 4(%esi,%edx,4),%xmm3 movhps 16(%esi,%edx,4),%xmm4 movhps 28(%esi,%edx,4),%xmm5 shufps $204,%xmm3,%xmm0 ## constant 11001100 shufps $204,%xmm4,%xmm1 ## constant 11001100 shufps $204,%xmm5,%xmm2 ## constant 11001100 movaps %xmm0,nb102_jzO(%esp) movaps %xmm1,nb102_jzH1(%esp) movaps %xmm2,nb102_jzH2(%esp) movaps nb102_ixO(%esp),%xmm0 movaps nb102_iyO(%esp),%xmm1 movaps nb102_izO(%esp),%xmm2 movaps nb102_ixO(%esp),%xmm3 movaps nb102_iyO(%esp),%xmm4 movaps nb102_izO(%esp),%xmm5 subps nb102_jxO(%esp),%xmm0 subps nb102_jyO(%esp),%xmm1 subps nb102_jzO(%esp),%xmm2 subps nb102_jxH1(%esp),%xmm3 subps nb102_jyH1(%esp),%xmm4 subps nb102_jzH1(%esp),%xmm5 movaps %xmm0,nb102_dxOO(%esp) movaps %xmm1,nb102_dyOO(%esp) movaps %xmm2,nb102_dzOO(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb102_dxOH1(%esp) movaps %xmm4,nb102_dyOH1(%esp) movaps %xmm5,nb102_dzOH1(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb102_rsqOO(%esp) movaps %xmm3,nb102_rsqOH1(%esp) movaps nb102_ixO(%esp),%xmm0 movaps nb102_iyO(%esp),%xmm1 movaps nb102_izO(%esp),%xmm2 movaps nb102_ixH1(%esp),%xmm3 movaps nb102_iyH1(%esp),%xmm4 movaps nb102_izH1(%esp),%xmm5 subps nb102_jxH2(%esp),%xmm0 subps nb102_jyH2(%esp),%xmm1 subps nb102_jzH2(%esp),%xmm2 subps nb102_jxO(%esp),%xmm3 subps nb102_jyO(%esp),%xmm4 subps nb102_jzO(%esp),%xmm5 movaps %xmm0,nb102_dxOH2(%esp) movaps %xmm1,nb102_dyOH2(%esp) movaps %xmm2,nb102_dzOH2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb102_dxH1O(%esp) movaps %xmm4,nb102_dyH1O(%esp) movaps %xmm5,nb102_dzH1O(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb102_rsqOH2(%esp) movaps %xmm3,nb102_rsqH1O(%esp) movaps nb102_ixH1(%esp),%xmm0 movaps nb102_iyH1(%esp),%xmm1 movaps nb102_izH1(%esp),%xmm2 movaps nb102_ixH1(%esp),%xmm3 movaps nb102_iyH1(%esp),%xmm4 movaps nb102_izH1(%esp),%xmm5 subps nb102_jxH1(%esp),%xmm0 subps nb102_jyH1(%esp),%xmm1 subps nb102_jzH1(%esp),%xmm2 subps nb102_jxH2(%esp),%xmm3 subps nb102_jyH2(%esp),%xmm4 subps nb102_jzH2(%esp),%xmm5 movaps %xmm0,nb102_dxH1H1(%esp) movaps %xmm1,nb102_dyH1H1(%esp) movaps %xmm2,nb102_dzH1H1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb102_dxH1H2(%esp) movaps %xmm4,nb102_dyH1H2(%esp) movaps %xmm5,nb102_dzH1H2(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb102_rsqH1H1(%esp) movaps %xmm3,nb102_rsqH1H2(%esp) movaps nb102_ixH2(%esp),%xmm0 movaps nb102_iyH2(%esp),%xmm1 movaps nb102_izH2(%esp),%xmm2 movaps nb102_ixH2(%esp),%xmm3 movaps nb102_iyH2(%esp),%xmm4 movaps nb102_izH2(%esp),%xmm5 subps nb102_jxO(%esp),%xmm0 subps nb102_jyO(%esp),%xmm1 subps nb102_jzO(%esp),%xmm2 subps nb102_jxH1(%esp),%xmm3 subps nb102_jyH1(%esp),%xmm4 subps nb102_jzH1(%esp),%xmm5 movaps %xmm0,nb102_dxH2O(%esp) movaps %xmm1,nb102_dyH2O(%esp) movaps %xmm2,nb102_dzH2O(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb102_dxH2H1(%esp) movaps %xmm4,nb102_dyH2H1(%esp) movaps %xmm5,nb102_dzH2H1(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm3,%xmm4 addps %xmm5,%xmm4 movaps %xmm0,nb102_rsqH2O(%esp) movaps %xmm4,nb102_rsqH2H1(%esp) movaps nb102_ixH2(%esp),%xmm0 movaps nb102_iyH2(%esp),%xmm1 movaps nb102_izH2(%esp),%xmm2 subps nb102_jxH2(%esp),%xmm0 subps nb102_jyH2(%esp),%xmm1 subps nb102_jzH2(%esp),%xmm2 movaps %xmm0,nb102_dxH2H2(%esp) movaps %xmm1,nb102_dyH2H2(%esp) movaps %xmm2,nb102_dzH2H2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 movaps %xmm0,nb102_rsqH2H2(%esp) ## start doing invsqrt use rsq values in xmm0, xmm4 rsqrtps %xmm0,%xmm1 rsqrtps %xmm4,%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb102_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps %xmm0,%xmm1 mulps %xmm4,%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb102_half(%esp),%xmm3 ## rinvH2H2 mulps nb102_half(%esp),%xmm7 ## rinvH2H1 movaps %xmm3,nb102_rinvH2H2(%esp) movaps %xmm7,nb102_rinvH2H1(%esp) rsqrtps nb102_rsqOO(%esp),%xmm1 rsqrtps nb102_rsqOH1(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb102_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb102_rsqOO(%esp),%xmm1 mulps nb102_rsqOH1(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb102_half(%esp),%xmm3 mulps nb102_half(%esp),%xmm7 movaps %xmm3,nb102_rinvOO(%esp) movaps %xmm7,nb102_rinvOH1(%esp) rsqrtps nb102_rsqOH2(%esp),%xmm1 rsqrtps nb102_rsqH1O(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb102_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb102_rsqOH2(%esp),%xmm1 mulps nb102_rsqH1O(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb102_half(%esp),%xmm3 mulps nb102_half(%esp),%xmm7 movaps %xmm3,nb102_rinvOH2(%esp) movaps %xmm7,nb102_rinvH1O(%esp) rsqrtps nb102_rsqH1H1(%esp),%xmm1 rsqrtps nb102_rsqH1H2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb102_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb102_rsqH1H1(%esp),%xmm1 mulps nb102_rsqH1H2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb102_half(%esp),%xmm3 mulps nb102_half(%esp),%xmm7 movaps %xmm3,nb102_rinvH1H1(%esp) movaps %xmm7,nb102_rinvH1H2(%esp) rsqrtps nb102_rsqH2O(%esp),%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb102_three(%esp),%xmm3 mulps nb102_rsqH2O(%esp),%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb102_half(%esp),%xmm3 movaps %xmm3,nb102_rinvH2O(%esp) ## start with OO interaction movaps nb102_rinvOO(%esp),%xmm0 movaps %xmm0,%xmm7 mulps %xmm0,%xmm0 mulps nb102_qqOO(%esp),%xmm7 mulps %xmm7,%xmm0 addps nb102_vctot(%esp),%xmm7 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb102_dxOO(%esp),%xmm0 mulps nb102_dyOO(%esp),%xmm1 mulps nb102_dzOO(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb102_fixO(%esp),%xmm0 addps nb102_fiyO(%esp),%xmm1 addps nb102_fizO(%esp),%xmm2 movaps %xmm3,nb102_fjxO(%esp) movaps %xmm4,nb102_fjyO(%esp) movaps %xmm5,nb102_fjzO(%esp) movaps %xmm0,nb102_fixO(%esp) movaps %xmm1,nb102_fiyO(%esp) movaps %xmm2,nb102_fizO(%esp) ## O-H1 interaction movaps nb102_rinvOH1(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb102_qqOH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fsOH1 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb102_dxOH1(%esp),%xmm0 mulps nb102_dyOH1(%esp),%xmm1 mulps nb102_dzOH1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb102_fixO(%esp),%xmm0 addps nb102_fiyO(%esp),%xmm1 addps nb102_fizO(%esp),%xmm2 movaps %xmm3,nb102_fjxH1(%esp) movaps %xmm4,nb102_fjyH1(%esp) movaps %xmm5,nb102_fjzH1(%esp) movaps %xmm0,nb102_fixO(%esp) movaps %xmm1,nb102_fiyO(%esp) movaps %xmm2,nb102_fizO(%esp) ## O-H2 interaction movaps nb102_rinvOH2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb102_qqOH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fsOH2 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb102_dxOH2(%esp),%xmm0 mulps nb102_dyOH2(%esp),%xmm1 mulps nb102_dzOH2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb102_fixO(%esp),%xmm0 addps nb102_fiyO(%esp),%xmm1 addps nb102_fizO(%esp),%xmm2 movaps %xmm3,nb102_fjxH2(%esp) movaps %xmm4,nb102_fjyH2(%esp) movaps %xmm5,nb102_fjzH2(%esp) movaps %xmm0,nb102_fixO(%esp) movaps %xmm1,nb102_fiyO(%esp) movaps %xmm2,nb102_fizO(%esp) ## H1-O interaction movaps nb102_rinvH1O(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb102_qqOH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fsH1O addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb102_fjxO(%esp),%xmm3 movaps nb102_fjyO(%esp),%xmm4 movaps nb102_fjzO(%esp),%xmm5 mulps nb102_dxH1O(%esp),%xmm0 mulps nb102_dyH1O(%esp),%xmm1 mulps nb102_dzH1O(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb102_fixH1(%esp),%xmm0 addps nb102_fiyH1(%esp),%xmm1 addps nb102_fizH1(%esp),%xmm2 movaps %xmm3,nb102_fjxO(%esp) movaps %xmm4,nb102_fjyO(%esp) movaps %xmm5,nb102_fjzO(%esp) movaps %xmm0,nb102_fixH1(%esp) movaps %xmm1,nb102_fiyH1(%esp) movaps %xmm2,nb102_fizH1(%esp) ## H1-H1 interaction movaps nb102_rinvH1H1(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb102_qqHH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fsH1H1 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb102_fjxH1(%esp),%xmm3 movaps nb102_fjyH1(%esp),%xmm4 movaps nb102_fjzH1(%esp),%xmm5 mulps nb102_dxH1H1(%esp),%xmm0 mulps nb102_dyH1H1(%esp),%xmm1 mulps nb102_dzH1H1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb102_fixH1(%esp),%xmm0 addps nb102_fiyH1(%esp),%xmm1 addps nb102_fizH1(%esp),%xmm2 movaps %xmm3,nb102_fjxH1(%esp) movaps %xmm4,nb102_fjyH1(%esp)
⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?