nb_kernel304_ia32_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 1,976 行 · 第 1/5 页

S
1,976 行
字号
12345
        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  nb304_two(%esp),%xmm7            ## two*Heps2         movaps nb304_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  nb304_vctot(%esp),%xmm5        movaps %xmm5,nb304_vctot(%esp)        xorps  %xmm1,%xmm1        mulps  nb304_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 nb304_dxH1H2(%esp),%xmm0        mulps nb304_dyH1H2(%esp),%xmm1        mulps nb304_dzH1H2(%esp),%xmm2        subps %xmm0,%xmm3        subps %xmm1,%xmm4        subps %xmm2,%xmm5        addps nb304_fixH1(%esp),%xmm0        addps nb304_fiyH1(%esp),%xmm1        addps nb304_fizH1(%esp),%xmm2        movaps %xmm3,nb304_fjxH2(%esp)        movaps %xmm4,nb304_fjyH2(%esp)        movaps %xmm5,nb304_fjzH2(%esp)        movaps %xmm0,nb304_fixH1(%esp)        movaps %xmm1,nb304_fiyH1(%esp)        movaps %xmm2,nb304_fizH1(%esp)        ## H1-M interaction          movaps nb304_rinvH1M(%esp),%xmm0        movaps %xmm0,%xmm1        mulps  nb304_rsqH1M(%esp),%xmm1   ## xmm1=r         mulps  nb304_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        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  nb304_two(%esp),%xmm7            ## two*Heps2         movaps nb304_qqMH(%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  nb304_vctot(%esp),%xmm5        movaps %xmm5,nb304_vctot(%esp)        xorps  %xmm1,%xmm1        mulps  nb304_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 nb304_dxH1M(%esp),%xmm0        mulps nb304_dyH1M(%esp),%xmm1        mulps nb304_dzH1M(%esp),%xmm2        subps %xmm0,%xmm3        subps %xmm1,%xmm4        subps %xmm2,%xmm5        addps nb304_fixH1(%esp),%xmm0        addps nb304_fiyH1(%esp),%xmm1        addps nb304_fizH1(%esp),%xmm2        movaps %xmm3,nb304_fjxM(%esp)        movaps %xmm4,nb304_fjyM(%esp)        movaps %xmm5,nb304_fjzM(%esp)        movaps %xmm0,nb304_fixH1(%esp)        movaps %xmm1,nb304_fiyH1(%esp)        movaps %xmm2,nb304_fizH1(%esp)        ## H2-H1 interaction         movaps nb304_rinvH2H1(%esp),%xmm0        movaps %xmm0,%xmm1        mulps  nb304_rsqH2H1(%esp),%xmm1   ## xmm1=r         mulps  nb304_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        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  nb304_two(%esp),%xmm7            ## two*Heps2         movaps nb304_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  nb304_vctot(%esp),%xmm5        movaps %xmm5,nb304_vctot(%esp)        xorps  %xmm1,%xmm1        mulps  nb304_tsc(%esp),%xmm3        mulps  %xmm0,%xmm3        subps  %xmm3,%xmm1        movaps %xmm1,%xmm0        movaps %xmm1,%xmm2        movaps nb304_fjxH1(%esp),%xmm3        movaps nb304_fjyH1(%esp),%xmm4        movaps nb304_fjzH1(%esp),%xmm5        mulps nb304_dxH2H1(%esp),%xmm0        mulps nb304_dyH2H1(%esp),%xmm1        mulps nb304_dzH2H1(%esp),%xmm2        subps %xmm0,%xmm3        subps %xmm1,%xmm4        subps %xmm2,%xmm5        addps nb304_fixH2(%esp),%xmm0        addps nb304_fiyH2(%esp),%xmm1        addps nb304_fizH2(%esp),%xmm2        movaps %xmm3,nb304_fjxH1(%esp)        movaps %xmm4,nb304_fjyH1(%esp)        movaps %xmm5,nb304_fjzH1(%esp)        movaps %xmm0,nb304_fixH2(%esp)        movaps %xmm1,nb304_fiyH2(%esp)        movaps %xmm2,nb304_fizH2(%esp)        ## H2-H2 interaction         movaps nb304_rinvH2H2(%esp),%xmm0        movaps %xmm0,%xmm1        mulps  nb304_rsqH2H2(%esp),%xmm1   ## xmm1=r         mulps  nb304_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        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  nb304_two(%esp),%xmm7            ## two*Heps2         movaps nb304_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  nb304_vctot(%esp),%xmm5        movaps %xmm5,nb304_vctot(%esp)        xorps  %xmm1,%xmm1        mulps  nb304_tsc(%esp),%xmm3        mulps  %xmm0,%xmm3        subps  %xmm3,%xmm1        movaps %xmm1,%xmm0        movaps %xmm1,%xmm2        movaps nb304_fjxH2(%esp),%xmm3        movaps nb304_fjyH2(%esp),%xmm4        movaps nb304_fjzH2(%esp),%xmm5        mulps nb304_dxH2H2(%esp),%xmm0        mulps nb304_dyH2H2(%esp),%xmm1        mulps nb304_dzH2H2(%esp),%xmm2        subps %xmm0,%xmm3        subps %xmm1,%xmm4        subps %xmm2,%xmm5        addps nb304_fixH2(%esp),%xmm0        addps nb304_fiyH2(%esp),%xmm1        addps nb304_fizH2(%esp),%xmm2        movaps %xmm3,nb304_fjxH2(%esp)        movaps %xmm4,nb304_fjyH2(%esp)        movaps %xmm5,nb304_fjzH2(%esp)        movaps %xmm0,nb304_fixH2(%esp)        movaps %xmm1,nb304_fiyH2(%esp)        movaps %xmm2,nb304_fizH2(%esp)        ## H2-M interaction         movaps nb304_rinvH2M(%esp),%xmm0        movaps %xmm0,%xmm1        mulps  nb304_rsqH2M(%esp),%xmm1   ## xmm1=r         mulps  nb304_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        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  nb304_two(%esp),%xmm7            ## two*Heps2         movaps nb304_qqMH(%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  nb304_vctot(%esp),%xmm5        movaps %xmm5,nb304_vctot(%esp)        xorps  %xmm1,%xmm1        mulps  nb304_tsc(%esp),%xmm3        mulps  %xmm0,%xmm3        subps  %xmm3,%xmm1        movaps %xmm1,%xmm0        movaps %xmm1,%xmm2        movaps nb304_fjxM(%esp),%xmm3        movaps nb304_fjyM(%esp),%xmm4        movaps nb304_fjzM(%esp),%xmm5        mulps nb304_dxH2M(%esp),%xmm0        mulps nb304_dyH2M(%esp),%xmm1        mulps nb304_dzH2M(%esp),%xmm2        subps %xmm0,%xmm3        subps %xmm1,%xmm4        subps %xmm2,%xmm5        addps nb304_fixH2(%esp),%xmm0        addps nb304_fiyH2(%esp),%xmm1        addps nb304_fizH2(%esp),%xmm2        movaps %xmm3,nb304_fjxM(%esp)        movaps %xmm4,nb304_fjyM(%esp)        movaps %xmm5,nb304_fjzM(%esp)        movaps %xmm0,nb304_fixH2(%esp)        movaps %xmm1,nb304_fiyH2(%esp)        movaps %xmm2,nb304_fizH2(%esp)        ## M-H1 interaction         movaps nb304_rinvMH1(%esp),%xmm0        movaps %xmm0,%xmm1        mulps  nb304_rsqMH1(%esp),%xmm1   ## xmm1=r         mulps  nb304_tsc(%esp),%xmm1        movhlps %xmm1,%xmm2
12345

nb_kernel304_ia32_sse.s - 源码说明

本页面展示了「最著名最快的分子模拟软件」中的 nb_kernel304_ia32_sse.s 源码文件，采用 S 编程语言编写，共 1,976 行代码。您可以在线阅读完整代码内容，也可以返回资源详情页下载完整源码包进行本地学习和开发。

虫虫开发者社区收录了大量与分子模拟相关的技术资源，包括源代码、技术文档、电路图等，是电子工程师和嵌入式开发者的专业学习平台。

⌨️ 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?