nb_kernel212_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,955 行 · 第 1/5 页
S
1,955 行
movl nb212_pos(%ebp),%esi ## base of pos[] leal (%eax,%eax,2),%eax ## replace jnr with j3 leal (%ebx,%ebx,2),%ebx leal (%ecx,%ecx,2),%ecx ## replace jnr with j3 leal (%edx,%edx,2),%edx ## move j coordinates to local temp variables movlps (%esi,%eax,4),%xmm2 movlps 12(%esi,%eax,4),%xmm3 movlps 24(%esi,%eax,4),%xmm4 movlps (%esi,%ebx,4),%xmm5 movlps 12(%esi,%ebx,4),%xmm6 movlps 24(%esi,%ebx,4),%xmm7 movhps (%esi,%ecx,4),%xmm2 movhps 12(%esi,%ecx,4),%xmm3 movhps 24(%esi,%ecx,4),%xmm4 movhps (%esi,%edx,4),%xmm5 movhps 12(%esi,%edx,4),%xmm6 movhps 24(%esi,%edx,4),%xmm7 ## current state: ## xmm2= jxOa jyOa jxOc jyOc ## xmm3= jxH1a jyH1a jxH1c jyH1c ## 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,nb212_jxO(%esp) movhlps %xmm6,%xmm2 ## xmm2= jyOa jyOb jyOc jyOd movaps %xmm2,nb212_jyO(%esp) movlhps %xmm3,%xmm1 movaps %xmm1,nb212_jxH1(%esp) movhlps %xmm7,%xmm3 movaps %xmm4,%xmm6 movaps %xmm3,nb212_jyH1(%esp) movlhps %xmm5,%xmm4 movaps %xmm4,nb212_jxH2(%esp) movhlps %xmm6,%xmm5 movaps %xmm5,nb212_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,nb212_jzO(%esp) movaps %xmm1,nb212_jzH1(%esp) movaps %xmm2,nb212_jzH2(%esp) movaps nb212_ixO(%esp),%xmm0 movaps nb212_iyO(%esp),%xmm1 movaps nb212_izO(%esp),%xmm2 movaps nb212_ixO(%esp),%xmm3 movaps nb212_iyO(%esp),%xmm4 movaps nb212_izO(%esp),%xmm5 subps nb212_jxO(%esp),%xmm0 subps nb212_jyO(%esp),%xmm1 subps nb212_jzO(%esp),%xmm2 subps nb212_jxH1(%esp),%xmm3 subps nb212_jyH1(%esp),%xmm4 subps nb212_jzH1(%esp),%xmm5 movaps %xmm0,nb212_dxOO(%esp) movaps %xmm1,nb212_dyOO(%esp) movaps %xmm2,nb212_dzOO(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb212_dxOH1(%esp) movaps %xmm4,nb212_dyOH1(%esp) movaps %xmm5,nb212_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,nb212_rsqOO(%esp) movaps %xmm3,nb212_rsqOH1(%esp) movaps nb212_ixO(%esp),%xmm0 movaps nb212_iyO(%esp),%xmm1 movaps nb212_izO(%esp),%xmm2 movaps nb212_ixH1(%esp),%xmm3 movaps nb212_iyH1(%esp),%xmm4 movaps nb212_izH1(%esp),%xmm5 subps nb212_jxH2(%esp),%xmm0 subps nb212_jyH2(%esp),%xmm1 subps nb212_jzH2(%esp),%xmm2 subps nb212_jxO(%esp),%xmm3 subps nb212_jyO(%esp),%xmm4 subps nb212_jzO(%esp),%xmm5 movaps %xmm0,nb212_dxOH2(%esp) movaps %xmm1,nb212_dyOH2(%esp) movaps %xmm2,nb212_dzOH2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb212_dxH1O(%esp) movaps %xmm4,nb212_dyH1O(%esp) movaps %xmm5,nb212_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,nb212_rsqOH2(%esp) movaps %xmm3,nb212_rsqH1O(%esp) movaps nb212_ixH1(%esp),%xmm0 movaps nb212_iyH1(%esp),%xmm1 movaps nb212_izH1(%esp),%xmm2 movaps nb212_ixH1(%esp),%xmm3 movaps nb212_iyH1(%esp),%xmm4 movaps nb212_izH1(%esp),%xmm5 subps nb212_jxH1(%esp),%xmm0 subps nb212_jyH1(%esp),%xmm1 subps nb212_jzH1(%esp),%xmm2 subps nb212_jxH2(%esp),%xmm3 subps nb212_jyH2(%esp),%xmm4 subps nb212_jzH2(%esp),%xmm5 movaps %xmm0,nb212_dxH1H1(%esp) movaps %xmm1,nb212_dyH1H1(%esp) movaps %xmm2,nb212_dzH1H1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb212_dxH1H2(%esp) movaps %xmm4,nb212_dyH1H2(%esp) movaps %xmm5,nb212_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,nb212_rsqH1H1(%esp) movaps %xmm3,nb212_rsqH1H2(%esp) movaps nb212_ixH2(%esp),%xmm0 movaps nb212_iyH2(%esp),%xmm1 movaps nb212_izH2(%esp),%xmm2 movaps nb212_ixH2(%esp),%xmm3 movaps nb212_iyH2(%esp),%xmm4 movaps nb212_izH2(%esp),%xmm5 subps nb212_jxO(%esp),%xmm0 subps nb212_jyO(%esp),%xmm1 subps nb212_jzO(%esp),%xmm2 subps nb212_jxH1(%esp),%xmm3 subps nb212_jyH1(%esp),%xmm4 subps nb212_jzH1(%esp),%xmm5 movaps %xmm0,nb212_dxH2O(%esp) movaps %xmm1,nb212_dyH2O(%esp) movaps %xmm2,nb212_dzH2O(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb212_dxH2H1(%esp) movaps %xmm4,nb212_dyH2H1(%esp) movaps %xmm5,nb212_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,nb212_rsqH2O(%esp) movaps %xmm4,nb212_rsqH2H1(%esp) movaps nb212_ixH2(%esp),%xmm0 movaps nb212_iyH2(%esp),%xmm1 movaps nb212_izH2(%esp),%xmm2 subps nb212_jxH2(%esp),%xmm0 subps nb212_jyH2(%esp),%xmm1 subps nb212_jzH2(%esp),%xmm2 movaps %xmm0,nb212_dxH2H2(%esp) movaps %xmm1,nb212_dyH2H2(%esp) movaps %xmm2,nb212_dzH2H2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 movaps %xmm0,nb212_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 nb212_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 nb212_half(%esp),%xmm3 ## rinvH2H2 mulps nb212_half(%esp),%xmm7 ## rinvH2H1 movaps %xmm3,nb212_rinvH2H2(%esp) movaps %xmm7,nb212_rinvH2H1(%esp) rsqrtps nb212_rsqOO(%esp),%xmm1 rsqrtps nb212_rsqOH1(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb212_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb212_rsqOO(%esp),%xmm1 mulps nb212_rsqOH1(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb212_half(%esp),%xmm3 mulps nb212_half(%esp),%xmm7 movaps %xmm3,nb212_rinvOO(%esp) movaps %xmm7,nb212_rinvOH1(%esp) rsqrtps nb212_rsqOH2(%esp),%xmm1 rsqrtps nb212_rsqH1O(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb212_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb212_rsqOH2(%esp),%xmm1 mulps nb212_rsqH1O(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb212_half(%esp),%xmm3 mulps nb212_half(%esp),%xmm7 movaps %xmm3,nb212_rinvOH2(%esp) movaps %xmm7,nb212_rinvH1O(%esp) rsqrtps nb212_rsqH1H1(%esp),%xmm1 rsqrtps nb212_rsqH1H2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb212_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb212_rsqH1H1(%esp),%xmm1 mulps nb212_rsqH1H2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb212_half(%esp),%xmm3 mulps nb212_half(%esp),%xmm7 movaps %xmm3,nb212_rinvH1H1(%esp) movaps %xmm7,nb212_rinvH1H2(%esp) rsqrtps nb212_rsqH2O(%esp),%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb212_three(%esp),%xmm3 mulps nb212_rsqH2O(%esp),%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb212_half(%esp),%xmm3 movaps %xmm3,nb212_rinvH2O(%esp) ## start with OO interaction movaps nb212_rinvOO(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb212_krf(%esp),%xmm5 mulps %xmm0,%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm1 mulps %xmm0,%xmm1 ## xmm1=rinvsix mulps nb212_rsqOO(%esp),%xmm5 ## xmm5=krsq movaps %xmm5,%xmm6 addps %xmm7,%xmm6 ## xmm6=rinv+ krsq subps nb212_crf(%esp),%xmm6 mulps nb212_qqOO(%esp),%xmm6 ## xmm6=voul=qq*(rinv+ krsq-crf) mulps nb212_two(%esp),%xmm5 subps %xmm5,%xmm7 ## xmm7=rinv-2*krsq mulps nb212_qqOO(%esp),%xmm7 ## xmm7 = coul part of fscal movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps nb212_c6(%esp),%xmm1 mulps nb212_c12(%esp),%xmm2 movaps %xmm2,%xmm3 subps %xmm1,%xmm3 ## xmm3=Vvdw12-Vvdw6 addps nb212_Vvdwtot(%esp),%xmm3 mulps nb212_six(%esp),%xmm1 mulps nb212_twelve(%esp),%xmm2 movaps %xmm3,nb212_Vvdwtot(%esp) subps %xmm1,%xmm2 addps %xmm7,%xmm2 addps nb212_vctot(%esp),%xmm6 ## local vctot summation variable mulps %xmm2,%xmm0 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb212_dxOO(%esp),%xmm0 mulps nb212_dyOO(%esp),%xmm1 mulps nb212_dzOO(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb212_fixO(%esp),%xmm0 addps nb212_fiyO(%esp),%xmm1 addps nb212_fizO(%esp),%xmm2 movaps %xmm3,nb212_fjxO(%esp) movaps %xmm4,nb212_fjyO(%esp) movaps %xmm5,nb212_fjzO(%esp) movaps %xmm0,nb212_fixO(%esp) movaps %xmm1,nb212_fiyO(%esp) movaps %xmm2,nb212_fizO(%esp) ## O-H1 interaction movaps nb212_rinvOH1(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb212_krf(%esp),%xmm5 movaps %xmm0,%xmm1 mulps nb212_rsqOH1(%esp),%xmm5 ## xmm5=krsq movaps %xmm5,%xmm4 addps %xmm7,%xmm4 ## xmm4=rinv+ krsq mulps %xmm0,%xmm0 subps nb212_crf(%esp),%xmm4 mulps nb212_qqOH(%esp),%xmm4 ## xmm4=voul=qq*(rinv+ krsq) mulps nb212_two(%esp),%xmm5 subps %xmm5,%xmm7 ## xmm7=rinv-2*krsq mulps nb212_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 nb212_dxOH1(%esp),%xmm0 mulps nb212_dyOH1(%esp),%xmm1 mulps nb212_dzOH1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5
⌨️ 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?