nb_kernel202_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,957 行 · 第 1/5 页
S
1,957 行
## xmm4= jxH2a jyH2a jxH2c jyH2c ## xmm5= jxOb jyOb jxOd jyOd ## xmm6= jxH1b jyH1b jxH1d jyH1d ## xmm7= jxH2b jyH2b jxH2d jyH2d movaps %xmm2,%xmm0 movaps %xmm3,%xmm1 unpcklps %xmm5,%xmm0 ## xmm0= jxOa jxOb jyOa jyOb unpcklps %xmm6,%xmm1 ## xmm1= jxH1a jxH1b jyH1a jyH1b unpckhps %xmm5,%xmm2 ## xmm2= jxOc jxOd jyOc jyOd unpckhps %xmm6,%xmm3 ## xmm3= jxH1c jxH1d jyH1c jyH1d movaps %xmm4,%xmm5 movaps %xmm0,%xmm6 unpcklps %xmm7,%xmm4 ## xmm4= jxH2a jxH2b jyH2a jyH2b unpckhps %xmm7,%xmm5 ## xmm5= jxH2c jxH2d jyH2c jyH2d movaps %xmm1,%xmm7 movlhps %xmm2,%xmm0 ## xmm0= jxOa jxOb jxOc jxOd movaps %xmm0,nb202_jxO(%esp) movhlps %xmm6,%xmm2 ## xmm2= jyOa jyOb jyOc jyOd movaps %xmm2,nb202_jyO(%esp) movlhps %xmm3,%xmm1 movaps %xmm1,nb202_jxH1(%esp) movhlps %xmm7,%xmm3 movaps %xmm4,%xmm6 movaps %xmm3,nb202_jyH1(%esp) movlhps %xmm5,%xmm4 movaps %xmm4,nb202_jxH2(%esp) movhlps %xmm6,%xmm5 movaps %xmm5,nb202_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,nb202_jzO(%esp) movaps %xmm1,nb202_jzH1(%esp) movaps %xmm2,nb202_jzH2(%esp) movaps nb202_ixO(%esp),%xmm0 movaps nb202_iyO(%esp),%xmm1 movaps nb202_izO(%esp),%xmm2 movaps nb202_ixO(%esp),%xmm3 movaps nb202_iyO(%esp),%xmm4 movaps nb202_izO(%esp),%xmm5 subps nb202_jxO(%esp),%xmm0 subps nb202_jyO(%esp),%xmm1 subps nb202_jzO(%esp),%xmm2 subps nb202_jxH1(%esp),%xmm3 subps nb202_jyH1(%esp),%xmm4 subps nb202_jzH1(%esp),%xmm5 movaps %xmm0,nb202_dxOO(%esp) movaps %xmm1,nb202_dyOO(%esp) movaps %xmm2,nb202_dzOO(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb202_dxOH1(%esp) movaps %xmm4,nb202_dyOH1(%esp) movaps %xmm5,nb202_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,nb202_rsqOO(%esp) movaps %xmm3,nb202_rsqOH1(%esp) movaps nb202_ixO(%esp),%xmm0 movaps nb202_iyO(%esp),%xmm1 movaps nb202_izO(%esp),%xmm2 movaps nb202_ixH1(%esp),%xmm3 movaps nb202_iyH1(%esp),%xmm4 movaps nb202_izH1(%esp),%xmm5 subps nb202_jxH2(%esp),%xmm0 subps nb202_jyH2(%esp),%xmm1 subps nb202_jzH2(%esp),%xmm2 subps nb202_jxO(%esp),%xmm3 subps nb202_jyO(%esp),%xmm4 subps nb202_jzO(%esp),%xmm5 movaps %xmm0,nb202_dxOH2(%esp) movaps %xmm1,nb202_dyOH2(%esp) movaps %xmm2,nb202_dzOH2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb202_dxH1O(%esp) movaps %xmm4,nb202_dyH1O(%esp) movaps %xmm5,nb202_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,nb202_rsqOH2(%esp) movaps %xmm3,nb202_rsqH1O(%esp) movaps nb202_ixH1(%esp),%xmm0 movaps nb202_iyH1(%esp),%xmm1 movaps nb202_izH1(%esp),%xmm2 movaps nb202_ixH1(%esp),%xmm3 movaps nb202_iyH1(%esp),%xmm4 movaps nb202_izH1(%esp),%xmm5 subps nb202_jxH1(%esp),%xmm0 subps nb202_jyH1(%esp),%xmm1 subps nb202_jzH1(%esp),%xmm2 subps nb202_jxH2(%esp),%xmm3 subps nb202_jyH2(%esp),%xmm4 subps nb202_jzH2(%esp),%xmm5 movaps %xmm0,nb202_dxH1H1(%esp) movaps %xmm1,nb202_dyH1H1(%esp) movaps %xmm2,nb202_dzH1H1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb202_dxH1H2(%esp) movaps %xmm4,nb202_dyH1H2(%esp) movaps %xmm5,nb202_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,nb202_rsqH1H1(%esp) movaps %xmm3,nb202_rsqH1H2(%esp) movaps nb202_ixH2(%esp),%xmm0 movaps nb202_iyH2(%esp),%xmm1 movaps nb202_izH2(%esp),%xmm2 movaps nb202_ixH2(%esp),%xmm3 movaps nb202_iyH2(%esp),%xmm4 movaps nb202_izH2(%esp),%xmm5 subps nb202_jxO(%esp),%xmm0 subps nb202_jyO(%esp),%xmm1 subps nb202_jzO(%esp),%xmm2 subps nb202_jxH1(%esp),%xmm3 subps nb202_jyH1(%esp),%xmm4 subps nb202_jzH1(%esp),%xmm5 movaps %xmm0,nb202_dxH2O(%esp) movaps %xmm1,nb202_dyH2O(%esp) movaps %xmm2,nb202_dzH2O(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb202_dxH2H1(%esp) movaps %xmm4,nb202_dyH2H1(%esp) movaps %xmm5,nb202_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,nb202_rsqH2O(%esp) movaps %xmm4,nb202_rsqH2H1(%esp) movaps nb202_ixH2(%esp),%xmm0 movaps nb202_iyH2(%esp),%xmm1 movaps nb202_izH2(%esp),%xmm2 subps nb202_jxH2(%esp),%xmm0 subps nb202_jyH2(%esp),%xmm1 subps nb202_jzH2(%esp),%xmm2 movaps %xmm0,nb202_dxH2H2(%esp) movaps %xmm1,nb202_dyH2H2(%esp) movaps %xmm2,nb202_dzH2H2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 movaps %xmm0,nb202_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 nb202_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 nb202_half(%esp),%xmm3 ## rinvH2H2 mulps nb202_half(%esp),%xmm7 ## rinvH2H1 movaps %xmm3,nb202_rinvH2H2(%esp) movaps %xmm7,nb202_rinvH2H1(%esp) rsqrtps nb202_rsqOO(%esp),%xmm1 rsqrtps nb202_rsqOH1(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb202_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb202_rsqOO(%esp),%xmm1 mulps nb202_rsqOH1(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb202_half(%esp),%xmm3 mulps nb202_half(%esp),%xmm7 movaps %xmm3,nb202_rinvOO(%esp) movaps %xmm7,nb202_rinvOH1(%esp) rsqrtps nb202_rsqOH2(%esp),%xmm1 rsqrtps nb202_rsqH1O(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb202_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb202_rsqOH2(%esp),%xmm1 mulps nb202_rsqH1O(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb202_half(%esp),%xmm3 mulps nb202_half(%esp),%xmm7 movaps %xmm3,nb202_rinvOH2(%esp) movaps %xmm7,nb202_rinvH1O(%esp) rsqrtps nb202_rsqH1H1(%esp),%xmm1 rsqrtps nb202_rsqH1H2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb202_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb202_rsqH1H1(%esp),%xmm1 mulps nb202_rsqH1H2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb202_half(%esp),%xmm3 mulps nb202_half(%esp),%xmm7 movaps %xmm3,nb202_rinvH1H1(%esp) movaps %xmm7,nb202_rinvH1H2(%esp) rsqrtps nb202_rsqH2O(%esp),%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb202_three(%esp),%xmm3 mulps nb202_rsqH2O(%esp),%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb202_half(%esp),%xmm3 movaps %xmm3,nb202_rinvH2O(%esp) ## start with OO interaction movaps nb202_rinvOO(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb202_krf(%esp),%xmm5 mulps %xmm0,%xmm0 ## xmm0=rinvsq mulps nb202_rsqOO(%esp),%xmm5 ## xmm5=krsq movaps %xmm5,%xmm6 addps %xmm7,%xmm6 ## xmm6=rinv+ krsq subps nb202_crf(%esp),%xmm6 mulps nb202_qqOO(%esp),%xmm6 ## xmm6=voul=qq*(rinv+ krsq-crf) mulps nb202_two(%esp),%xmm5 subps %xmm5,%xmm7 ## xmm7=rinv-2*krsq mulps nb202_qqOO(%esp),%xmm7 ## xmm7 = coul part of fscal addps nb202_vctot(%esp),%xmm6 ## local vctot summation variable mulps %xmm7,%xmm0 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb202_dxOO(%esp),%xmm0 mulps nb202_dyOO(%esp),%xmm1 mulps nb202_dzOO(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb202_fixO(%esp),%xmm0 addps nb202_fiyO(%esp),%xmm1 addps nb202_fizO(%esp),%xmm2 movaps %xmm3,nb202_fjxO(%esp) movaps %xmm4,nb202_fjyO(%esp) movaps %xmm5,nb202_fjzO(%esp) movaps %xmm0,nb202_fixO(%esp) movaps %xmm1,nb202_fiyO(%esp) movaps %xmm2,nb202_fizO(%esp) ## O-H1 interaction movaps nb202_rinvOH1(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb202_krf(%esp),%xmm5 movaps %xmm0,%xmm1 mulps nb202_rsqOH1(%esp),%xmm5 ## xmm5=krsq movaps %xmm5,%xmm4 addps %xmm7,%xmm4 ## xmm4=r inv+ krsq subps nb202_crf(%esp),%xmm4 mulps %xmm0,%xmm0 mulps nb202_qqOH(%esp),%xmm4 ## xmm4=voul=qq*(rinv+ krsq-crf) mulps nb202_two(%esp),%xmm5 subps %xmm5,%xmm7 ## xmm7=rinv-2*krsq mulps nb202_qqOH(%esp),%xmm7 ## xmm7 = coul part of fscal addps %xmm4,%xmm6 ## add to local vctot mulps %xmm7,%xmm0 ## fsOH1 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb202_dxOH1(%esp),%xmm0 mulps nb202_dyOH1(%esp),%xmm1 mulps nb202_dzOH1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb202_fixO(%esp),%xmm0 addps nb202_fiyO(%esp),%xmm1 addps nb202_fizO(%esp),%xmm2 movaps %xmm3,nb202_fjxH1(%esp) movaps %xmm4,nb202_fjyH1(%esp) movaps %xmm5,nb202_fjzH1(%esp) movaps %xmm0,nb202_fixO(%esp) movaps %xmm1,nb202_fiyO(%esp) movaps %xmm2,nb202_fizO(%esp) ## O-H2 interaction movaps nb202_rinvOH2(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb202_krf(%esp),%xmm5 movaps %xmm0,%xmm1 mulps nb202_rsqOH2(%esp),%xmm5 ## xmm5=krsq movaps %xmm5,%xmm4 addps %xmm7,%xmm4 ## xmm4=r inv+ krsq subps nb202_crf(%esp),%xmm4 mulps %xmm0,%xmm0 mulps nb202_qqOH(%esp),%xmm4 ## xmm4=voul=qq*(rinv+ krsq-crf) mulps nb202_two(%esp),%xmm5 subps %xmm5,%xmm7 ## xmm7=rinv-2*krsq mulps nb202_qqOH(%esp),%xmm7 ## xmm7 = coul part of fscal addps %xmm4,%xmm6 ## add to local vctot mulps %xmm7,%xmm0 ## fsOH2 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb202_dxOH2(%esp),%xmm0 mulps nb202_dyOH2(%esp),%xmm1 mulps nb202_dzOH2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb202_fixO(%esp),%xmm0 addps nb202_fiyO(%esp),%xmm1 addps nb202_fizO(%esp),%xmm2 movaps %xmm3,nb202_fjxH2(%esp) movaps %xmm4,nb202_fjyH2(%esp) movaps %xmm5,nb202_fjzH2(%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?