nb_kernel234_ia32_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 1,980 行 · 第 1/5 页

S
1,980 行
字号
12345
        movl  nb234_jjnr(%ebp),%eax        shll  $2,%ecx        addl  %ecx,%eax        movl  %eax,nb234_innerjjnr(%esp)        ## pointer to jjnr[nj0]         movl  %edx,%ecx        subl  $4,%edx        addl  nb234_ninner(%esp),%ecx        movl  %ecx,nb234_ninner(%esp)        addl  $0,%edx        movl  %edx,nb234_innerk(%esp)   ## number of innerloop atoms         jge   _nb_kernel234_ia32_sse.nb234_unroll_loop        jmp   _nb_kernel234_ia32_sse.nb234_single_check_nb_kernel234_ia32_sse.nb234_unroll_loop:         ## quad-unroll innerloop here         movl  nb234_innerjjnr(%esp),%edx        ## pointer to jjnr[k]         movl  (%edx),%eax        movl  4(%edx),%ebx        movl  8(%edx),%ecx        movl  12(%edx),%edx             ## eax-edx=jnr1-4         addl $16,nb234_innerjjnr(%esp)             ## advance pointer (unroll 4)         movl nb234_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        ## Load Ox, Oy, Oz, H1x         movlps (%esi,%eax,4),%xmm1      ##  Oxa   Oya    -    -        movlps (%esi,%ecx,4),%xmm4      ##  Oxc   Oyc    -    -        movhps (%esi,%ebx,4),%xmm1      ##  Oxa   Oya   Oxb   Oyb         movhps (%esi,%edx,4),%xmm4      ##  Oxc   Oyc   Oxd   Oyd         movaps %xmm1,%xmm0              ##  Oxa   Oya   Oxb   Oyb         shufps $0x88,%xmm4,%xmm0       ##  Oxa   Oxb   Oxc   Oxd        shufps $0xDD,%xmm4,%xmm1       ##  Oya   Oyb   Oyc   Oyd        movlps 8(%esi,%eax,4),%xmm3     ##  Oza  H1xa    -    -        movlps 8(%esi,%ecx,4),%xmm5     ##  Ozc  H1xc    -    -        movhps 8(%esi,%ebx,4),%xmm3     ##  Oza  H1xa   Ozb  H1xb         movhps 8(%esi,%edx,4),%xmm5     ##  Ozc  H1xc   Ozd  H1xd         movaps %xmm3,%xmm2              ##  Oza  H1xa   Ozb  H1xb         shufps $0x88,%xmm5,%xmm2       ##  Oza   Ozb   Ozc   Ozd        shufps $0xDD,%xmm5,%xmm3       ## H1xa  H1xb  H1xc  H1xd        ## coordinates in xmm0-xmm3             ## store        movaps %xmm0,nb234_jxO(%esp)        movaps %xmm1,nb234_jyO(%esp)        movaps %xmm2,nb234_jzO(%esp)        movaps %xmm3,nb234_jxH1(%esp)        ## Load H1y H1z H2x H2y         movlps 16(%esi,%eax,4),%xmm1        movlps 16(%esi,%ecx,4),%xmm4        movhps 16(%esi,%ebx,4),%xmm1        movhps 16(%esi,%edx,4),%xmm4        movaps %xmm1,%xmm0        shufps $0x88,%xmm4,%xmm0        shufps $0xDD,%xmm4,%xmm1        movlps 24(%esi,%eax,4),%xmm3        movlps 24(%esi,%ecx,4),%xmm5        movhps 24(%esi,%ebx,4),%xmm3        movhps 24(%esi,%edx,4),%xmm5        movaps %xmm3,%xmm2        shufps $0x88,%xmm5,%xmm2        shufps $0xDD,%xmm5,%xmm3        ## coordinates in xmm0-xmm3             ## store        movaps %xmm0,nb234_jyH1(%esp)        movaps %xmm1,nb234_jzH1(%esp)        movaps %xmm2,nb234_jxH2(%esp)        movaps %xmm3,nb234_jyH2(%esp)        ## Load H2z Mx My Mz         movlps 32(%esi,%eax,4),%xmm1        movlps 32(%esi,%ecx,4),%xmm4        movhps 32(%esi,%ebx,4),%xmm1        movhps 32(%esi,%edx,4),%xmm4        movaps %xmm1,%xmm0        shufps $0x88,%xmm4,%xmm0        shufps $0xDD,%xmm4,%xmm1        movlps 40(%esi,%eax,4),%xmm3        movlps 40(%esi,%ecx,4),%xmm5        movhps 40(%esi,%ebx,4),%xmm3        movhps 40(%esi,%edx,4),%xmm5        movaps %xmm3,%xmm2        shufps $0x88,%xmm5,%xmm2        shufps $0xDD,%xmm5,%xmm3        ## coordinates in xmm0-xmm3             ## store        movaps %xmm0,nb234_jzH2(%esp)        movaps %xmm1,nb234_jxM(%esp)        movaps %xmm2,nb234_jyM(%esp)        movaps %xmm3,nb234_jzM(%esp)        ## start calculating pairwise distances        movaps nb234_ixO(%esp),%xmm0        movaps nb234_iyO(%esp),%xmm1        movaps nb234_izO(%esp),%xmm2        movaps nb234_ixH1(%esp),%xmm3        movaps nb234_iyH1(%esp),%xmm4        movaps nb234_izH1(%esp),%xmm5        subps  nb234_jxO(%esp),%xmm0        subps  nb234_jyO(%esp),%xmm1        subps  nb234_jzO(%esp),%xmm2        subps  nb234_jxH1(%esp),%xmm3        subps  nb234_jyH1(%esp),%xmm4        subps  nb234_jzH1(%esp),%xmm5        movaps %xmm0,nb234_dxOO(%esp)        movaps %xmm1,nb234_dyOO(%esp)        movaps %xmm2,nb234_dzOO(%esp)        mulps  %xmm0,%xmm0        mulps  %xmm1,%xmm1        mulps  %xmm2,%xmm2        movaps %xmm3,nb234_dxH1H1(%esp)        movaps %xmm4,nb234_dyH1H1(%esp)        movaps %xmm5,nb234_dzH1H1(%esp)        mulps  %xmm3,%xmm3        mulps  %xmm4,%xmm4        mulps  %xmm5,%xmm5        addps  %xmm1,%xmm0        addps  %xmm2,%xmm0        addps  %xmm4,%xmm3        addps  %xmm5,%xmm3        movaps %xmm0,nb234_rsqOO(%esp)        movaps %xmm3,nb234_rsqH1H1(%esp)        movaps nb234_ixH1(%esp),%xmm0        movaps nb234_iyH1(%esp),%xmm1        movaps nb234_izH1(%esp),%xmm2        movaps %xmm0,%xmm3        movaps %xmm1,%xmm4        movaps %xmm2,%xmm5        subps  nb234_jxH2(%esp),%xmm0        subps  nb234_jyH2(%esp),%xmm1        subps  nb234_jzH2(%esp),%xmm2        subps  nb234_jxM(%esp),%xmm3        subps  nb234_jyM(%esp),%xmm4        subps  nb234_jzM(%esp),%xmm5        movaps %xmm0,nb234_dxH1H2(%esp)        movaps %xmm1,nb234_dyH1H2(%esp)        movaps %xmm2,nb234_dzH1H2(%esp)        mulps  %xmm0,%xmm0        mulps  %xmm1,%xmm1        mulps  %xmm2,%xmm2        movaps %xmm3,nb234_dxH1M(%esp)        movaps %xmm4,nb234_dyH1M(%esp)        movaps %xmm5,nb234_dzH1M(%esp)        mulps  %xmm3,%xmm3        mulps  %xmm4,%xmm4        mulps  %xmm5,%xmm5        addps  %xmm1,%xmm0        addps  %xmm2,%xmm0        addps  %xmm4,%xmm3        addps  %xmm5,%xmm3        movaps %xmm0,nb234_rsqH1H2(%esp)        movaps %xmm3,nb234_rsqH1M(%esp)        movaps nb234_ixH2(%esp),%xmm0        movaps nb234_iyH2(%esp),%xmm1        movaps nb234_izH2(%esp),%xmm2        movaps %xmm0,%xmm3        movaps %xmm1,%xmm4        movaps %xmm2,%xmm5        subps  nb234_jxH1(%esp),%xmm0        subps  nb234_jyH1(%esp),%xmm1        subps  nb234_jzH1(%esp),%xmm2        subps  nb234_jxH2(%esp),%xmm3        subps  nb234_jyH2(%esp),%xmm4        subps  nb234_jzH2(%esp),%xmm5        movaps %xmm0,nb234_dxH2H1(%esp)        movaps %xmm1,nb234_dyH2H1(%esp)        movaps %xmm2,nb234_dzH2H1(%esp)        mulps  %xmm0,%xmm0        mulps  %xmm1,%xmm1        mulps  %xmm2,%xmm2        movaps %xmm3,nb234_dxH2H2(%esp)        movaps %xmm4,nb234_dyH2H2(%esp)        movaps %xmm5,nb234_dzH2H2(%esp)        mulps  %xmm3,%xmm3        mulps  %xmm4,%xmm4        mulps  %xmm5,%xmm5        addps  %xmm1,%xmm0        addps  %xmm2,%xmm0        addps  %xmm4,%xmm3        addps  %xmm5,%xmm3        movaps %xmm0,nb234_rsqH2H1(%esp)        movaps %xmm3,nb234_rsqH2H2(%esp)        movaps nb234_ixH2(%esp),%xmm0        movaps nb234_iyH2(%esp),%xmm1        movaps nb234_izH2(%esp),%xmm2        movaps nb234_ixM(%esp),%xmm3        movaps nb234_iyM(%esp),%xmm4        movaps nb234_izM(%esp),%xmm5        subps  nb234_jxM(%esp),%xmm0        subps  nb234_jyM(%esp),%xmm1        subps  nb234_jzM(%esp),%xmm2        subps  nb234_jxH1(%esp),%xmm3        subps  nb234_jyH1(%esp),%xmm4        subps  nb234_jzH1(%esp),%xmm5        movaps %xmm0,nb234_dxH2M(%esp)        movaps %xmm1,nb234_dyH2M(%esp)        movaps %xmm2,nb234_dzH2M(%esp)        mulps  %xmm0,%xmm0        mulps  %xmm1,%xmm1        mulps  %xmm2,%xmm2        movaps %xmm3,nb234_dxMH1(%esp)        movaps %xmm4,nb234_dyMH1(%esp)        movaps %xmm5,nb234_dzMH1(%esp)        mulps  %xmm3,%xmm3        mulps  %xmm4,%xmm4        mulps  %xmm5,%xmm5        addps  %xmm1,%xmm0        addps  %xmm2,%xmm0        addps  %xmm3,%xmm4        addps  %xmm5,%xmm4        movaps %xmm0,nb234_rsqH2M(%esp)        movaps %xmm4,nb234_rsqMH1(%esp)        movaps nb234_ixM(%esp),%xmm0        movaps nb234_iyM(%esp),%xmm1        movaps nb234_izM(%esp),%xmm2        movaps %xmm0,%xmm3        movaps %xmm1,%xmm4        movaps %xmm2,%xmm5        subps  nb234_jxH2(%esp),%xmm0        subps  nb234_jyH2(%esp),%xmm1        subps  nb234_jzH2(%esp),%xmm2        subps  nb234_jxM(%esp),%xmm3        subps  nb234_jyM(%esp),%xmm4        subps  nb234_jzM(%esp),%xmm5        movaps %xmm0,nb234_dxMH2(%esp)        movaps %xmm1,nb234_dyMH2(%esp)        movaps %xmm2,nb234_dzMH2(%esp)        mulps  %xmm0,%xmm0        mulps  %xmm1,%xmm1        mulps  %xmm2,%xmm2        movaps %xmm3,nb234_dxMM(%esp)        movaps %xmm4,nb234_dyMM(%esp)        movaps %xmm5,nb234_dzMM(%esp)        mulps  %xmm3,%xmm3        mulps  %xmm4,%xmm4        mulps  %xmm5,%xmm5        addps  %xmm1,%xmm0        addps  %xmm2,%xmm0        addps  %xmm3,%xmm4        addps  %xmm5,%xmm4        movaps %xmm0,nb234_rsqMH2(%esp)        movaps %xmm4,nb234_rsqMM(%esp)        ## start by doing invsqrt for OO        rsqrtps nb234_rsqOO(%esp),%xmm1        movaps  %xmm1,%xmm2        mulps   %xmm1,%xmm1        movaps  nb234_three(%esp),%xmm3        mulps   nb234_rsqOO(%esp),%xmm1        subps   %xmm1,%xmm3        mulps   %xmm2,%xmm3        mulps   nb234_half(%esp),%xmm3        movaps  %xmm3,nb234_rinvOO(%esp)        ## more invsqrt ops - do two at a time.        rsqrtps nb234_rsqH1H1(%esp),%xmm1        rsqrtps nb234_rsqH1H2(%esp),%xmm5        movaps  %xmm1,%xmm2        movaps  %xmm5,%xmm6        mulps   %xmm1,%xmm1        mulps   %xmm5,%xmm5        movaps  nb234_three(%esp),%xmm3        movaps  %xmm3,%xmm7        mulps   nb234_rsqH1H1(%esp),%xmm1        mulps   nb234_rsqH1H2(%esp),%xmm5        subps   %xmm1,%xmm3        subps   %xmm5,%xmm7        mulps   %xmm2,%xmm3        mulps   %xmm6,%xmm7        mulps   nb234_half(%esp),%xmm3   ## rinvH1H1         mulps   nb234_half(%esp),%xmm7   ## rinvH1H2         movaps  %xmm3,nb234_rinvH1H1(%esp)        movaps  %xmm7,nb234_rinvH1H2(%esp)        rsqrtps nb234_rsqH1M(%esp),%xmm1        rsqrtps nb234_rsqH2H1(%esp),%xmm5        movaps  %xmm1,%xmm2        movaps  %xmm5,%xmm6        mulps   %xmm1,%xmm1        mulps   %xmm5,%xmm5        movaps  nb234_three(%esp),%xmm3        movaps  %xmm3,%xmm7        mulps   nb234_rsqH1M(%esp),%xmm1        mulps   nb234_rsqH2H1(%esp),%xmm5        subps   %xmm1,%xmm3        subps   %xmm5,%xmm7        mulps   %xmm2,%xmm3        mulps   %xmm6,%xmm7        mulps   nb234_half(%esp),%xmm3        mulps   nb234_half(%esp),%xmm7        movaps  %xmm3,nb234_rinvH1M(%esp)        movaps  %xmm7,nb234_rinvH2H1(%esp)        rsqrtps nb234_rsqH2H2(%esp),%xmm1        rsqrtps nb234_rsqH2M(%esp),%xmm5        movaps  %xmm1,%xmm2        movaps  %xmm5,%xmm6        mulps   %xmm1,%xmm1        mulps   %xmm5,%xmm5        movaps  nb234_three(%esp),%xmm3        movaps  %xmm3,%xmm7        mulps   nb234_rsqH2H2(%esp),%xmm1        mulps   nb234_rsqH2M(%esp),%xmm5        subps   %xmm1,%xmm3        subps   %xmm5,%xmm7        mulps   %xmm2,%xmm3        mulps   %xmm6,%xmm7        mulps   nb234_half(%esp),%xmm3        mulps   nb234_half(%esp),%xmm7        movaps  %xmm3,nb234_rinvH2H2(%esp)        movaps  %xmm7,nb234_rinvH2M(%esp)        rsqrtps nb234_rsqMH1(%esp),%xmm1        rsqrtps nb234_rsqMH2(%esp),%xmm5        movaps  %xmm1,%xmm2        movaps  %xmm5,%xmm6        mulps   %xmm1,%xmm1        mulps   %xmm5,%xmm5        movaps  nb234_three(%esp),%xmm3        movaps  %xmm3,%xmm7        mulps   nb234_rsqMH1(%esp),%xmm1        mulps   nb234_rsqMH2(%esp),%xmm5        subps   %xmm1,%xmm3        subps   %xmm5,%xmm7        mulps   %xmm2,%xmm3        mulps   %xmm6,%xmm7        mulps   nb234_half(%esp),%xmm3        mulps   nb234_half(%esp),%xmm7        movaps  %xmm3,nb234_rinvMH1(%esp)        movaps  %xmm7,nb234_rinvMH2(%esp)        rsqrtps nb234_rsqMM(%esp),%xmm1        movaps  %xmm1,%xmm2        mulps   %xmm1,%xmm1        movaps  nb234_three(%esp),%xmm3        mulps   nb234_rsqMM(%esp),%xmm1        subps   %xmm1,%xmm3        mulps   %xmm2,%xmm3        mulps   nb234_half(%esp),%xmm3        movaps  %xmm3,nb234_rinvMM(%esp)        ## start with OO LJ interaction        movaps nb234_rinvOO(%esp),%xmm0        movaps %xmm0,%xmm1        mulps  nb234_rsqOO(%esp),%xmm1   ## xmm1=r         mulps  nb234_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 $3,%mm6    pslld $3,%mm7    movd %eax,%mm0    movd %ebx,%mm1    movd %ecx,%mm2    movd %edx,%mm3    movd %mm6,%eax    psrlq $32,%mm6    movd %mm7,%ecx    psrlq $32,%mm7    movd %mm6,%ebx    movd %mm7,%edx    movl nb234_VFtab(%ebp),%esi    ## dispersion     movlps (%esi,%eax,4),%xmm5    movlps (%esi,%ecx,4),%xmm7    movhps (%esi,%ebx,4),%xmm5    movhps (%esi,%edx,4),%xmm7 ## got half 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
12345

nb_kernel234_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_kernel234_ia32_sse.s 源码文件，采用 S 编程语言编写，共 1,980 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?