nb_kernel332_ia32_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,043 行 · 第 1/5 页

S
2,043
字号
    addps  %xmm6,%xmm5    addps  %xmm7,%xmm5      ## xmm5=Fp     mulps  nb332_two(%esp),%xmm7         ## two*Heps2     addps  %xmm6,%xmm7    addps  %xmm5,%xmm7 ## xmm7=FF     mulps  %xmm1,%xmm5 ## xmm5=eps*Fp     addps  %xmm4,%xmm5 ## xmm5=VV     movaps nb332_c12(%esp),%xmm4    mulps  %xmm4,%xmm7 ## fijR     mulps  %xmm4,%xmm5 ## Vvdw12     addps  nb332_fstmp(%esp),%xmm7    addps  nb332_Vvdwtot(%esp),%xmm5    movaps %xmm5,nb332_Vvdwtot(%esp)    xorps  %xmm1,%xmm1    mulps nb332_tsc(%esp),%xmm7    mulps %xmm0,%xmm7    subps  %xmm7,%xmm1        movaps %xmm1,%xmm0        movaps %xmm1,%xmm2        xorps %xmm3,%xmm3        movaps %xmm3,%xmm4        movaps %xmm3,%xmm5        mulps nb332_dxOO(%esp),%xmm0        mulps nb332_dyOO(%esp),%xmm1        mulps nb332_dzOO(%esp),%xmm2        subps %xmm0,%xmm3        subps %xmm1,%xmm4        subps %xmm2,%xmm5        addps nb332_fixO(%esp),%xmm0        addps nb332_fiyO(%esp),%xmm1        addps nb332_fizO(%esp),%xmm2        movaps %xmm3,nb332_fjxO(%esp)        movaps %xmm4,nb332_fjyO(%esp)        movaps %xmm5,nb332_fjzO(%esp)        movaps %xmm0,nb332_fixO(%esp)        movaps %xmm1,nb332_fiyO(%esp)        movaps %xmm2,nb332_fizO(%esp)        ## O-H1 interaction         movaps nb332_rinvOH1(%esp),%xmm0        movaps %xmm0,%xmm1        mulps  nb332_rsqOH1(%esp),%xmm1   ## xmm1=r         mulps  nb332_tsc(%esp),%xmm1        movhlps %xmm1,%xmm2    cvttps2pi %xmm1,%mm6    cvttps2pi %xmm2,%mm7    ## mm6/mm7 contain lu indices     cvtpi2ps %mm6,%xmm3    cvtpi2ps %mm7,%xmm2        movlhps  %xmm2,%xmm3        subps    %xmm3,%xmm1    ## xmm1=eps     movaps %xmm1,%xmm2    mulps  %xmm2,%xmm2      ## xmm2=eps2     pslld $2,%mm6    pslld $2,%mm7    movd %mm6,%eax    psrlq $32,%mm6    movd %mm7,%ecx    psrlq $32,%mm7    movd %mm6,%ebx    movd %mm7,%edx    leal  (%eax,%eax,2),%eax    leal  (%ebx,%ebx,2),%ebx    leal  (%ecx,%ecx,2),%ecx    leal  (%edx,%edx,2),%edx    movlps (%esi,%eax,4),%xmm5    movlps (%esi,%ecx,4),%xmm7    movhps (%esi,%ebx,4),%xmm5    movhps (%esi,%edx,4),%xmm7 ## got half coulomb table     movaps %xmm5,%xmm4    shufps $136,%xmm7,%xmm4 ## constant 10001000    shufps $221,%xmm7,%xmm5 ## constant 11011101    movlps 8(%esi,%eax,4),%xmm7    movlps 8(%esi,%ecx,4),%xmm3    movhps 8(%esi,%ebx,4),%xmm7    movhps 8(%esi,%edx,4),%xmm3    ## other half of coulomb table      movaps %xmm7,%xmm6    shufps $136,%xmm3,%xmm6 ## constant 10001000    shufps $221,%xmm3,%xmm7 ## constant 11011101    ## coulomb table ready, in xmm4-xmm7      mulps  %xmm1,%xmm6      ## xmm6=Geps     mulps  %xmm2,%xmm7      ## xmm7=Heps2     addps  %xmm6,%xmm5    addps  %xmm7,%xmm5      ## xmm5=Fp     mulps  nb332_two(%esp),%xmm7         ## two*Heps2     movaps nb332_qqOH(%esp),%xmm3    addps  %xmm6,%xmm7    addps  %xmm5,%xmm7 ## xmm7=FF     mulps  %xmm1,%xmm5 ## xmm5=eps*Fp     addps  %xmm4,%xmm5 ## xmm5=VV     mulps  %xmm3,%xmm5 ## vcoul=qq*VV      mulps  %xmm7,%xmm3 ## fijC=FF*qq     ## at this point mm5 contains vcoul and mm3 fijC     addps  nb332_vctot(%esp),%xmm5    movaps %xmm5,nb332_vctot(%esp)        xorps  %xmm1,%xmm1        mulps  nb332_tsc(%esp),%xmm3        mulps  %xmm0,%xmm3        subps  %xmm3,%xmm1        movaps %xmm1,%xmm0        movaps %xmm1,%xmm2        xorps %xmm3,%xmm3        movaps %xmm3,%xmm4        movaps %xmm3,%xmm5        mulps nb332_dxOH1(%esp),%xmm0        mulps nb332_dyOH1(%esp),%xmm1        mulps nb332_dzOH1(%esp),%xmm2        subps %xmm0,%xmm3        subps %xmm1,%xmm4        subps %xmm2,%xmm5        addps nb332_fixO(%esp),%xmm0        addps nb332_fiyO(%esp),%xmm1        addps nb332_fizO(%esp),%xmm2        movaps %xmm3,nb332_fjxH1(%esp)        movaps %xmm4,nb332_fjyH1(%esp)        movaps %xmm5,nb332_fjzH1(%esp)        movaps %xmm0,nb332_fixO(%esp)        movaps %xmm1,nb332_fiyO(%esp)        movaps %xmm2,nb332_fizO(%esp)        ## O-H2 interaction          movaps nb332_rinvOH2(%esp),%xmm0        movaps %xmm0,%xmm1        mulps  nb332_rsqOH2(%esp),%xmm1   ## xmm1=r         mulps  nb332_tsc(%esp),%xmm1        movhlps %xmm1,%xmm2    cvttps2pi %xmm1,%mm6    cvttps2pi %xmm2,%mm7    ## mm6/mm7 contain lu indices     cvtpi2ps %mm6,%xmm3    cvtpi2ps %mm7,%xmm2        movlhps  %xmm2,%xmm3        subps    %xmm3,%xmm1    ## xmm1=eps     movaps %xmm1,%xmm2    mulps  %xmm2,%xmm2      ## xmm2=eps2     pslld $2,%mm6    pslld $2,%mm7    movd %mm6,%eax    psrlq $32,%mm6    movd %mm7,%ecx    psrlq $32,%mm7    movd %mm6,%ebx    movd %mm7,%edx    leal  (%eax,%eax,2),%eax    leal  (%ebx,%ebx,2),%ebx    leal  (%ecx,%ecx,2),%ecx    leal  (%edx,%edx,2),%edx    movlps (%esi,%eax,4),%xmm5    movlps (%esi,%ecx,4),%xmm7    movhps (%esi,%ebx,4),%xmm5    movhps (%esi,%edx,4),%xmm7 ## got half coulomb table     movaps %xmm5,%xmm4    shufps $136,%xmm7,%xmm4 ## constant 10001000    shufps $221,%xmm7,%xmm5 ## constant 11011101    movlps 8(%esi,%eax,4),%xmm7    movlps 8(%esi,%ecx,4),%xmm3    movhps 8(%esi,%ebx,4),%xmm7    movhps 8(%esi,%edx,4),%xmm3    ## other half of coulomb table      movaps %xmm7,%xmm6    shufps $136,%xmm3,%xmm6 ## constant 10001000    shufps $221,%xmm3,%xmm7 ## constant 11011101    ## coulomb table ready, in xmm4-xmm7      mulps  %xmm1,%xmm6      ## xmm6=Geps     mulps  %xmm2,%xmm7      ## xmm7=Heps2     addps  %xmm6,%xmm5    addps  %xmm7,%xmm5      ## xmm5=Fp     mulps  nb332_two(%esp),%xmm7         ## two*Heps2     movaps nb332_qqOH(%esp),%xmm3    addps  %xmm6,%xmm7    addps  %xmm5,%xmm7 ## xmm7=FF     mulps  %xmm1,%xmm5 ## xmm5=eps*Fp     addps  %xmm4,%xmm5 ## xmm5=VV     mulps  %xmm3,%xmm5 ## vcoul=qq*VV      mulps  %xmm7,%xmm3 ## fijC=FF*qq     ## at this point mm5 contains vcoul and mm3 fijC     addps  nb332_vctot(%esp),%xmm5    movaps %xmm5,nb332_vctot(%esp)        xorps  %xmm1,%xmm1        mulps  nb332_tsc(%esp),%xmm3        mulps  %xmm0,%xmm3        subps  %xmm3,%xmm1        movaps %xmm1,%xmm0        movaps %xmm1,%xmm2        xorps %xmm3,%xmm3        movaps %xmm3,%xmm4        movaps %xmm3,%xmm5        mulps nb332_dxOH2(%esp),%xmm0        mulps nb332_dyOH2(%esp),%xmm1        mulps nb332_dzOH2(%esp),%xmm2        subps %xmm0,%xmm3        subps %xmm1,%xmm4        subps %xmm2,%xmm5        addps nb332_fixO(%esp),%xmm0        addps nb332_fiyO(%esp),%xmm1        addps nb332_fizO(%esp),%xmm2        movaps %xmm3,nb332_fjxH2(%esp)        movaps %xmm4,nb332_fjyH2(%esp)        movaps %xmm5,nb332_fjzH2(%esp)        movaps %xmm0,nb332_fixO(%esp)        movaps %xmm1,nb332_fiyO(%esp)        movaps %xmm2,nb332_fizO(%esp)        ## H1-O interaction         movaps nb332_rinvH1O(%esp),%xmm0        movaps %xmm0,%xmm1        mulps  nb332_rsqH1O(%esp),%xmm1   ## xmm1=r         mulps  nb332_tsc(%esp),%xmm1        movhlps %xmm1,%xmm2    cvttps2pi %xmm1,%mm6    cvttps2pi %xmm2,%mm7    ## mm6/mm7 contain lu indices     cvtpi2ps %mm6,%xmm3    cvtpi2ps %mm7,%xmm2        movlhps  %xmm2,%xmm3        subps    %xmm3,%xmm1    ## xmm1=eps     movaps %xmm1,%xmm2    mulps  %xmm2,%xmm2      ## xmm2=eps2     pslld $2,%mm6    pslld $2,%mm7    movd %mm6,%eax    psrlq $32,%mm6    movd %mm7,%ecx    psrlq $32,%mm7    movd %mm6,%ebx    movd %mm7,%edx    leal  (%eax,%eax,2),%eax    leal  (%ebx,%ebx,2),%ebx    leal  (%ecx,%ecx,2),%ecx    leal  (%edx,%edx,2),%edx    movlps (%esi,%eax,4),%xmm5    movlps (%esi,%ecx,4),%xmm7    movhps (%esi,%ebx,4),%xmm5    movhps (%esi,%edx,4),%xmm7 ## got half coulomb table     movaps %xmm5,%xmm4    shufps $136,%xmm7,%xmm4 ## constant 10001000    shufps $221,%xmm7,%xmm5 ## constant 11011101    movlps 8(%esi,%eax,4),%xmm7    movlps 8(%esi,%ecx,4),%xmm3    movhps 8(%esi,%ebx,4),%xmm7    movhps 8(%esi,%edx,4),%xmm3    ## other half of coulomb table      movaps %xmm7,%xmm6    shufps $136,%xmm3,%xmm6 ## constant 10001000    shufps $221,%xmm3,%xmm7 ## constant 11011101    ## coulomb table ready, in xmm4-xmm7      mulps  %xmm1,%xmm6      ## xmm6=Geps     mulps  %xmm2,%xmm7      ## xmm7=Heps2     addps  %xmm6,%xmm5    addps  %xmm7,%xmm5      ## xmm5=Fp     mulps  nb332_two(%esp),%xmm7         ## two*Heps2     movaps nb332_qqOH(%esp),%xmm3    addps  %xmm6,%xmm7    addps  %xmm5,%xmm7 ## xmm7=FF     mulps  %xmm1,%xmm5 ## xmm5=eps*Fp     addps  %xmm4,%xmm5 ## xmm5=VV     mulps  %xmm3,%xmm5 ## vcoul=qq*VV      mulps  %xmm7,%xmm3 ## fijC=FF*qq     ## at this point mm5 contains vcoul and mm3 fijC     addps  nb332_vctot(%esp),%xmm5    movaps %xmm5,nb332_vctot(%esp)        xorps  %xmm1,%xmm1        mulps  nb332_tsc(%esp),%xmm3        mulps  %xmm0,%xmm3        subps  %xmm3,%xmm1        movaps %xmm1,%xmm0        movaps %xmm1,%xmm2        movaps nb332_fjxO(%esp),%xmm3        movaps nb332_fjyO(%esp),%xmm4        movaps nb332_fjzO(%esp),%xmm5        mulps nb332_dxH1O(%esp),%xmm0        mulps nb332_dyH1O(%esp),%xmm1        mulps nb332_dzH1O(%esp),%xmm2        subps %xmm0,%xmm3        subps %xmm1,%xmm4        subps %xmm2,%xmm5        addps nb332_fixH1(%esp),%xmm0        addps nb332_fiyH1(%esp),%xmm1        addps nb332_fizH1(%esp),%xmm2        movaps %xmm3,nb332_fjxO(%esp)        movaps %xmm4,nb332_fjyO(%esp)        movaps %xmm5,nb332_fjzO(%esp)        movaps %xmm0,nb332_fixH1(%esp)        movaps %xmm1,nb332_fiyH1(%esp)        movaps %xmm2,nb332_fizH1(%esp)        ## H1-H1 interaction         movaps nb332_rinvH1H1(%esp),%xmm0        movaps %xmm0,%xmm1        mulps  nb332_rsqH1H1(%esp),%xmm1   ## xmm1=r         mulps  nb332_tsc(%esp),%xmm1        movhlps %xmm1,%xmm2    cvttps2pi %xmm1,%mm6    cvttps2pi %xmm2,%mm7    ## mm6/mm7 contain lu indices     cvtpi2ps %mm6,%xmm3    cvtpi2ps %mm7,%xmm2        movlhps  %xmm2,%xmm3        subps    %xmm3,%xmm1    ## xmm1=eps     movaps %xmm1,%xmm2    mulps  %xmm2,%xmm2      ## xmm2=eps2     pslld $2,%mm6    pslld $2,%mm7    movd %mm6,%eax    psrlq $32,%mm6    movd %mm7,%ecx    psrlq $32,%mm7    movd %mm6,%ebx    movd %mm7,%edx    leal  (%eax,%eax,2),%eax    leal  (%ebx,%ebx,2),%ebx    leal  (%ecx,%ecx,2),%ecx    leal  (%edx,%edx,2),%edx    movlps (%esi,%eax,4),%xmm5    movlps (%esi,%ecx,4),%xmm7    movhps (%esi,%ebx,4),%xmm5    movhps (%esi,%edx,4),%xmm7 ## got half coulomb table     movaps %xmm5,%xmm4    shufps $136,%xmm7,%xmm4 ## constant 10001000    shufps $221,%xmm7,%xmm5 ## constant 11011101    movlps 8(%esi,%eax,4),%xmm7    movlps 8(%esi,%ecx,4),%xmm3    movhps 8(%esi,%ebx,4),%xmm7    movhps 8(%esi,%edx,4),%xmm3    ## other half of coulomb table      movaps %xmm7,%xmm6    shufps $136,%xmm3,%xmm6 ## constant 10001000    shufps $221,%xmm3,%xmm7 ## constant 11011101    ## coulomb table ready, in xmm4-xmm7      mulps  %xmm1,%xmm6      ## xmm6=Geps     mulps  %xmm2,%xmm7      ## xmm7=Heps2     addps  %xmm6,%xmm5    addps  %xmm7,%xmm5      ## xmm5=Fp     mulps  nb332_two(%esp),%xmm7         ## two*Heps2     movaps nb332_qqHH(%esp),%xmm3    addps  %xmm6,%xmm7    addps  %xmm5,%xmm7 ## xmm7=FF     mulps  %xmm1,%xmm5 ## xmm5=eps*Fp     addps  %xmm4,%xmm5 ## xmm5=VV     mulps  %xmm3,%xmm5 ## vcoul=qq*VV      mulps  %xmm7,%xmm3 ## fijC=FF*qq     ## at this point mm5 contains vcoul and mm3 fijC     addps  nb332_vctot(%esp),%xmm5    movaps %xmm5,nb332_vctot(%esp)        xorps  %xmm1,%xmm1        mulps  nb332_tsc(%esp),%xmm3        mulps  %xmm0,%xmm3        subps  %xmm3,%xmm1        movaps %xmm1,%xmm0        movaps %xmm1,%xmm2        movaps nb332_fjxH1(%esp),%xmm3        movaps nb332_fjyH1(%esp),%xmm4        movaps nb332_fjzH1(%esp),%xmm5        mulps nb332_dxH1H1(%esp),%xmm0        mulps nb332_dyH1H1(%esp),%xmm1        mulps nb332_dzH1H1(%esp),%xmm2        subps %xmm0,%xmm3        subps %xmm1,%xmm4        subps %xmm2,%xmm5        addps nb332_fixH1(%esp),%xmm0        addps nb332_fiyH1(%esp),%xmm1        addps nb332_fizH1(%esp),%xmm2        movaps %xmm3,nb332_fjxH1(%esp)        movaps %xmm4,nb332_fjyH1(%esp)        movaps %xmm5,nb332_fjzH1(%esp)        movaps %xmm0,nb332_fixH1(%esp)        movaps %xmm1,nb332_fiyH1(%esp)        movaps %xmm2,nb332_fizH1(%esp)        ## H1-H2 interaction         movaps nb332_rinvH1H2(%esp),%xmm0        movaps %xmm0,%xmm1        mulps  nb332_rsqH1H2(%esp),%xmm1   ## xmm1=r         mulps  nb332_tsc(%esp),%xmm1        movhlps %xmm1,%xmm2

⌨️ 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?