nb_kernel112_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,970 行 · 第 1/5 页
S
1,970 行
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,nb112_jxO(%esp) movhlps %xmm6,%xmm2 ## xmm2= jyOa jyOb jyOc jyOd movaps %xmm2,nb112_jyO(%esp) movlhps %xmm3,%xmm1 movaps %xmm1,nb112_jxH1(%esp) movhlps %xmm7,%xmm3 movaps %xmm4,%xmm6 movaps %xmm3,nb112_jyH1(%esp) movlhps %xmm5,%xmm4 movaps %xmm4,nb112_jxH2(%esp) movhlps %xmm6,%xmm5 movaps %xmm5,nb112_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,nb112_jzO(%esp) movaps %xmm1,nb112_jzH1(%esp) movaps %xmm2,nb112_jzH2(%esp) movaps nb112_ixO(%esp),%xmm0 movaps nb112_iyO(%esp),%xmm1 movaps nb112_izO(%esp),%xmm2 movaps nb112_ixO(%esp),%xmm3 movaps nb112_iyO(%esp),%xmm4 movaps nb112_izO(%esp),%xmm5 subps nb112_jxO(%esp),%xmm0 subps nb112_jyO(%esp),%xmm1 subps nb112_jzO(%esp),%xmm2 subps nb112_jxH1(%esp),%xmm3 subps nb112_jyH1(%esp),%xmm4 subps nb112_jzH1(%esp),%xmm5 movaps %xmm0,nb112_dxOO(%esp) movaps %xmm1,nb112_dyOO(%esp) movaps %xmm2,nb112_dzOO(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb112_dxOH1(%esp) movaps %xmm4,nb112_dyOH1(%esp) movaps %xmm5,nb112_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,nb112_rsqOO(%esp) movaps %xmm3,nb112_rsqOH1(%esp) movaps nb112_ixO(%esp),%xmm0 movaps nb112_iyO(%esp),%xmm1 movaps nb112_izO(%esp),%xmm2 movaps nb112_ixH1(%esp),%xmm3 movaps nb112_iyH1(%esp),%xmm4 movaps nb112_izH1(%esp),%xmm5 subps nb112_jxH2(%esp),%xmm0 subps nb112_jyH2(%esp),%xmm1 subps nb112_jzH2(%esp),%xmm2 subps nb112_jxO(%esp),%xmm3 subps nb112_jyO(%esp),%xmm4 subps nb112_jzO(%esp),%xmm5 movaps %xmm0,nb112_dxOH2(%esp) movaps %xmm1,nb112_dyOH2(%esp) movaps %xmm2,nb112_dzOH2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb112_dxH1O(%esp) movaps %xmm4,nb112_dyH1O(%esp) movaps %xmm5,nb112_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,nb112_rsqOH2(%esp) movaps %xmm3,nb112_rsqH1O(%esp) movaps nb112_ixH1(%esp),%xmm0 movaps nb112_iyH1(%esp),%xmm1 movaps nb112_izH1(%esp),%xmm2 movaps nb112_ixH1(%esp),%xmm3 movaps nb112_iyH1(%esp),%xmm4 movaps nb112_izH1(%esp),%xmm5 subps nb112_jxH1(%esp),%xmm0 subps nb112_jyH1(%esp),%xmm1 subps nb112_jzH1(%esp),%xmm2 subps nb112_jxH2(%esp),%xmm3 subps nb112_jyH2(%esp),%xmm4 subps nb112_jzH2(%esp),%xmm5 movaps %xmm0,nb112_dxH1H1(%esp) movaps %xmm1,nb112_dyH1H1(%esp) movaps %xmm2,nb112_dzH1H1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb112_dxH1H2(%esp) movaps %xmm4,nb112_dyH1H2(%esp) movaps %xmm5,nb112_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,nb112_rsqH1H1(%esp) movaps %xmm3,nb112_rsqH1H2(%esp) movaps nb112_ixH2(%esp),%xmm0 movaps nb112_iyH2(%esp),%xmm1 movaps nb112_izH2(%esp),%xmm2 movaps nb112_ixH2(%esp),%xmm3 movaps nb112_iyH2(%esp),%xmm4 movaps nb112_izH2(%esp),%xmm5 subps nb112_jxO(%esp),%xmm0 subps nb112_jyO(%esp),%xmm1 subps nb112_jzO(%esp),%xmm2 subps nb112_jxH1(%esp),%xmm3 subps nb112_jyH1(%esp),%xmm4 subps nb112_jzH1(%esp),%xmm5 movaps %xmm0,nb112_dxH2O(%esp) movaps %xmm1,nb112_dyH2O(%esp) movaps %xmm2,nb112_dzH2O(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb112_dxH2H1(%esp) movaps %xmm4,nb112_dyH2H1(%esp) movaps %xmm5,nb112_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,nb112_rsqH2O(%esp) movaps %xmm4,nb112_rsqH2H1(%esp) movaps nb112_ixH2(%esp),%xmm0 movaps nb112_iyH2(%esp),%xmm1 movaps nb112_izH2(%esp),%xmm2 subps nb112_jxH2(%esp),%xmm0 subps nb112_jyH2(%esp),%xmm1 subps nb112_jzH2(%esp),%xmm2 movaps %xmm0,nb112_dxH2H2(%esp) movaps %xmm1,nb112_dyH2H2(%esp) movaps %xmm2,nb112_dzH2H2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 movaps %xmm0,nb112_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 nb112_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 nb112_half(%esp),%xmm3 ## rinvH2H2 mulps nb112_half(%esp),%xmm7 ## rinvH2H1 movaps %xmm3,nb112_rinvH2H2(%esp) movaps %xmm7,nb112_rinvH2H1(%esp) rsqrtps nb112_rsqOO(%esp),%xmm1 rsqrtps nb112_rsqOH1(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb112_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb112_rsqOO(%esp),%xmm1 mulps nb112_rsqOH1(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb112_half(%esp),%xmm3 mulps nb112_half(%esp),%xmm7 movaps %xmm3,nb112_rinvOO(%esp) movaps %xmm7,nb112_rinvOH1(%esp) rsqrtps nb112_rsqOH2(%esp),%xmm1 rsqrtps nb112_rsqH1O(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb112_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb112_rsqOH2(%esp),%xmm1 mulps nb112_rsqH1O(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb112_half(%esp),%xmm3 mulps nb112_half(%esp),%xmm7 movaps %xmm3,nb112_rinvOH2(%esp) movaps %xmm7,nb112_rinvH1O(%esp) rsqrtps nb112_rsqH1H1(%esp),%xmm1 rsqrtps nb112_rsqH1H2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb112_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb112_rsqH1H1(%esp),%xmm1 mulps nb112_rsqH1H2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb112_half(%esp),%xmm3 mulps nb112_half(%esp),%xmm7 movaps %xmm3,nb112_rinvH1H1(%esp) movaps %xmm7,nb112_rinvH1H2(%esp) rsqrtps nb112_rsqH2O(%esp),%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb112_three(%esp),%xmm3 mulps nb112_rsqH2O(%esp),%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb112_half(%esp),%xmm3 movaps %xmm3,nb112_rinvH2O(%esp) ## start with OO interaction movaps nb112_rinvOO(%esp),%xmm0 movaps %xmm0,%xmm7 mulps %xmm0,%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm1 mulps %xmm0,%xmm1 ## xmm1=rinvsix mulps nb112_qqOO(%esp),%xmm7 movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps nb112_c6(%esp),%xmm1 mulps nb112_c12(%esp),%xmm2 movaps %xmm2,%xmm3 subps %xmm1,%xmm3 ## xmm3=Vvdw12-Vvdw6 addps nb112_Vvdwtot(%esp),%xmm3 mulps nb112_six(%esp),%xmm1 mulps nb112_twelve(%esp),%xmm2 movaps %xmm3,nb112_Vvdwtot(%esp) subps %xmm1,%xmm2 addps %xmm7,%xmm2 addps nb112_vctot(%esp),%xmm7 mulps %xmm2,%xmm0 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb112_dxOO(%esp),%xmm0 mulps nb112_dyOO(%esp),%xmm1 mulps nb112_dzOO(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb112_fixO(%esp),%xmm0 addps nb112_fiyO(%esp),%xmm1 addps nb112_fizO(%esp),%xmm2 movaps %xmm3,nb112_fjxO(%esp) movaps %xmm4,nb112_fjyO(%esp) movaps %xmm5,nb112_fjzO(%esp) movaps %xmm0,nb112_fixO(%esp) movaps %xmm1,nb112_fiyO(%esp) movaps %xmm2,nb112_fizO(%esp) ## O-H1 interaction movaps nb112_rinvOH1(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb112_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 nb112_dxOH1(%esp),%xmm0 mulps nb112_dyOH1(%esp),%xmm1 mulps nb112_dzOH1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb112_fixO(%esp),%xmm0 addps nb112_fiyO(%esp),%xmm1 addps nb112_fizO(%esp),%xmm2 movaps %xmm3,nb112_fjxH1(%esp) movaps %xmm4,nb112_fjyH1(%esp) movaps %xmm5,nb112_fjzH1(%esp) movaps %xmm0,nb112_fixO(%esp) movaps %xmm1,nb112_fiyO(%esp) movaps %xmm2,nb112_fizO(%esp) ## O-H2 interaction movaps nb112_rinvOH2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb112_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 nb112_dxOH2(%esp),%xmm0 mulps nb112_dyOH2(%esp),%xmm1 mulps nb112_dzOH2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb112_fixO(%esp),%xmm0 addps nb112_fiyO(%esp),%xmm1 addps nb112_fizO(%esp),%xmm2 movaps %xmm3,nb112_fjxH2(%esp) movaps %xmm4,nb112_fjyH2(%esp) movaps %xmm5,nb112_fjzH2(%esp) movaps %xmm0,nb112_fixO(%esp) movaps %xmm1,nb112_fiyO(%esp) movaps %xmm2,nb112_fizO(%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?