nb_kernel334_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,029 行 · 第 1/5 页
S
2,029 行
movl nb334_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,nb334_jxO(%esp) movaps %xmm1,nb334_jyO(%esp) movaps %xmm2,nb334_jzO(%esp) movaps %xmm3,nb334_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,nb334_jyH1(%esp) movaps %xmm1,nb334_jzH1(%esp) movaps %xmm2,nb334_jxH2(%esp) movaps %xmm3,nb334_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,nb334_jzH2(%esp) movaps %xmm1,nb334_jxM(%esp) movaps %xmm2,nb334_jyM(%esp) movaps %xmm3,nb334_jzM(%esp) ## start calculating pairwise distances movaps nb334_ixO(%esp),%xmm0 movaps nb334_iyO(%esp),%xmm1 movaps nb334_izO(%esp),%xmm2 movaps nb334_ixH1(%esp),%xmm3 movaps nb334_iyH1(%esp),%xmm4 movaps nb334_izH1(%esp),%xmm5 subps nb334_jxO(%esp),%xmm0 subps nb334_jyO(%esp),%xmm1 subps nb334_jzO(%esp),%xmm2 subps nb334_jxH1(%esp),%xmm3 subps nb334_jyH1(%esp),%xmm4 subps nb334_jzH1(%esp),%xmm5 movaps %xmm0,nb334_dxOO(%esp) movaps %xmm1,nb334_dyOO(%esp) movaps %xmm2,nb334_dzOO(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb334_dxH1H1(%esp) movaps %xmm4,nb334_dyH1H1(%esp) movaps %xmm5,nb334_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,nb334_rsqOO(%esp) movaps %xmm3,nb334_rsqH1H1(%esp) movaps nb334_ixH1(%esp),%xmm0 movaps nb334_iyH1(%esp),%xmm1 movaps nb334_izH1(%esp),%xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 subps nb334_jxH2(%esp),%xmm0 subps nb334_jyH2(%esp),%xmm1 subps nb334_jzH2(%esp),%xmm2 subps nb334_jxM(%esp),%xmm3 subps nb334_jyM(%esp),%xmm4 subps nb334_jzM(%esp),%xmm5 movaps %xmm0,nb334_dxH1H2(%esp) movaps %xmm1,nb334_dyH1H2(%esp) movaps %xmm2,nb334_dzH1H2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb334_dxH1M(%esp) movaps %xmm4,nb334_dyH1M(%esp) movaps %xmm5,nb334_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,nb334_rsqH1H2(%esp) movaps %xmm3,nb334_rsqH1M(%esp) movaps nb334_ixH2(%esp),%xmm0 movaps nb334_iyH2(%esp),%xmm1 movaps nb334_izH2(%esp),%xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 subps nb334_jxH1(%esp),%xmm0 subps nb334_jyH1(%esp),%xmm1 subps nb334_jzH1(%esp),%xmm2 subps nb334_jxH2(%esp),%xmm3 subps nb334_jyH2(%esp),%xmm4 subps nb334_jzH2(%esp),%xmm5 movaps %xmm0,nb334_dxH2H1(%esp) movaps %xmm1,nb334_dyH2H1(%esp) movaps %xmm2,nb334_dzH2H1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb334_dxH2H2(%esp) movaps %xmm4,nb334_dyH2H2(%esp) movaps %xmm5,nb334_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,nb334_rsqH2H1(%esp) movaps %xmm3,nb334_rsqH2H2(%esp) movaps nb334_ixH2(%esp),%xmm0 movaps nb334_iyH2(%esp),%xmm1 movaps nb334_izH2(%esp),%xmm2 movaps nb334_ixM(%esp),%xmm3 movaps nb334_iyM(%esp),%xmm4 movaps nb334_izM(%esp),%xmm5 subps nb334_jxM(%esp),%xmm0 subps nb334_jyM(%esp),%xmm1 subps nb334_jzM(%esp),%xmm2 subps nb334_jxH1(%esp),%xmm3 subps nb334_jyH1(%esp),%xmm4 subps nb334_jzH1(%esp),%xmm5 movaps %xmm0,nb334_dxH2M(%esp) movaps %xmm1,nb334_dyH2M(%esp) movaps %xmm2,nb334_dzH2M(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb334_dxMH1(%esp) movaps %xmm4,nb334_dyMH1(%esp) movaps %xmm5,nb334_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,nb334_rsqH2M(%esp) movaps %xmm4,nb334_rsqMH1(%esp) movaps nb334_ixM(%esp),%xmm0 movaps nb334_iyM(%esp),%xmm1 movaps nb334_izM(%esp),%xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 subps nb334_jxH2(%esp),%xmm0 subps nb334_jyH2(%esp),%xmm1 subps nb334_jzH2(%esp),%xmm2 subps nb334_jxM(%esp),%xmm3 subps nb334_jyM(%esp),%xmm4 subps nb334_jzM(%esp),%xmm5 movaps %xmm0,nb334_dxMH2(%esp) movaps %xmm1,nb334_dyMH2(%esp) movaps %xmm2,nb334_dzMH2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb334_dxMM(%esp) movaps %xmm4,nb334_dyMM(%esp) movaps %xmm5,nb334_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,nb334_rsqMH2(%esp) movaps %xmm4,nb334_rsqMM(%esp) ## Invsqrt for O-O rsqrtps nb334_rsqOO(%esp),%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb334_three(%esp),%xmm3 mulps nb334_rsqOO(%esp),%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb334_half(%esp),%xmm3 ## rinvOO movaps %xmm3,nb334_rinvOO(%esp) ## Invsqrt for H1-H1 and H1-H2 rsqrtps nb334_rsqH1H1(%esp),%xmm1 rsqrtps nb334_rsqH1H2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb334_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb334_rsqH1H1(%esp),%xmm1 mulps nb334_rsqH1H2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb334_half(%esp),%xmm3 ## rinvH1H1 mulps nb334_half(%esp),%xmm7 ## rinvH1H2 movaps %xmm3,nb334_rinvH1H1(%esp) movaps %xmm7,nb334_rinvH1H2(%esp) rsqrtps nb334_rsqH1M(%esp),%xmm1 rsqrtps nb334_rsqH2H1(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb334_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb334_rsqH1M(%esp),%xmm1 mulps nb334_rsqH2H1(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb334_half(%esp),%xmm3 mulps nb334_half(%esp),%xmm7 movaps %xmm3,nb334_rinvH1M(%esp) movaps %xmm7,nb334_rinvH2H1(%esp) rsqrtps nb334_rsqH2H2(%esp),%xmm1 rsqrtps nb334_rsqH2M(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb334_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb334_rsqH2H2(%esp),%xmm1 mulps nb334_rsqH2M(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb334_half(%esp),%xmm3 mulps nb334_half(%esp),%xmm7 movaps %xmm3,nb334_rinvH2H2(%esp) movaps %xmm7,nb334_rinvH2M(%esp) rsqrtps nb334_rsqMH1(%esp),%xmm1 rsqrtps nb334_rsqMH2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb334_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb334_rsqMH1(%esp),%xmm1 mulps nb334_rsqMH2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb334_half(%esp),%xmm3 mulps nb334_half(%esp),%xmm7 movaps %xmm3,nb334_rinvMH1(%esp) movaps %xmm7,nb334_rinvMH2(%esp) rsqrtps nb334_rsqMM(%esp),%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb334_three(%esp),%xmm3 mulps nb334_rsqMM(%esp),%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb334_half(%esp),%xmm3 movaps %xmm3,nb334_rinvMM(%esp) ## start with OO table interaction movaps nb334_rinvOO(%esp),%xmm0 movaps %xmm0,%xmm1 mulps nb334_rsqOO(%esp),%xmm1 ## xmm1=r mulps nb334_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 %eax,%mm0 movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movl nb334_VFtab(%ebp),%esi 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 ## load dispersion table data into xmm4-xmm7 movlps 16(%esi,%eax,4),%xmm5 movlps 16(%esi,%ecx,4),%xmm7 movhps 16(%esi,%ebx,4),%xmm5 movhps 16(%esi,%edx,4),%xmm7 ## got half table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 24(%esi,%eax,4),%xmm7 movlps 24(%esi,%ecx,4),%xmm3 movhps 24(%esi,%ebx,4),%xmm7 movhps 24(%esi,%edx,4),%xmm3 ## other half of table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## 10001000 shufps $221,%xmm3,%xmm7 ## 11011101 ## dispersion table YFGH ready in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 ## F+Geps addps %xmm7,%xmm5 ## xmm5=Fp=F+Geps+Heps2 mulps nb334_two(%esp),%xmm7 ## 2*Heps2 addps %xmm6,%xmm7 ## Geps+2*Heps2 addps %xmm5,%xmm7 ## xmm7=FF = Fp+Geps+2*Heps2 mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV = Y+eps*Fp movaps nb334_c6(%esp),%xmm4 mulps %xmm4,%xmm7 ## fijD mulps %xmm4,%xmm5 ## Vvdw6 ## put scalar force on stack Update Vvdwtot directly addps nb334_Vvdwtot(%esp),%xmm5 movaps %xmm7,nb334_fstmp(%esp) ## fscal movaps %xmm5,nb334_Vvdwtot(%esp) ## load repulsion table data into xmm4-xmm7 movlps 32(%esi,%eax,4),%xmm5 movlps 32(%esi,%ecx,4),%xmm7 movhps 32(%esi,%ebx,4),%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?