nb_kernel214_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,984 行 · 第 1/5 页
S
1,984 行
movl nb214_faction(%ebp),%edi movl nb214_jjnr(%ebp),%eax shll $2,%ecx addl %ecx,%eax movl %eax,nb214_innerjjnr(%esp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb214_ninner(%esp),%ecx movl %ecx,nb214_ninner(%esp) addl $0,%edx movl %edx,nb214_innerk(%esp) ## number of innerloop atoms jge _nb_kernel214_ia32_sse.nb214_unroll_loop jmp _nb_kernel214_ia32_sse.nb214_single_check_nb_kernel214_ia32_sse.nb214_unroll_loop: ## quad-unroll innerloop here movl nb214_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,nb214_innerjjnr(%esp) ## advance pointer (unroll 4) movl nb214_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,nb214_jxO(%esp) movaps %xmm1,nb214_jyO(%esp) movaps %xmm2,nb214_jzO(%esp) movaps %xmm3,nb214_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,nb214_jyH1(%esp) movaps %xmm1,nb214_jzH1(%esp) movaps %xmm2,nb214_jxH2(%esp) movaps %xmm3,nb214_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,nb214_jzH2(%esp) movaps %xmm1,nb214_jxM(%esp) movaps %xmm2,nb214_jyM(%esp) movaps %xmm3,nb214_jzM(%esp) ## start calculating pairwise distances movaps nb214_ixO(%esp),%xmm0 movaps nb214_iyO(%esp),%xmm1 movaps nb214_izO(%esp),%xmm2 movaps nb214_ixH1(%esp),%xmm3 movaps nb214_iyH1(%esp),%xmm4 movaps nb214_izH1(%esp),%xmm5 subps nb214_jxO(%esp),%xmm0 subps nb214_jyO(%esp),%xmm1 subps nb214_jzO(%esp),%xmm2 subps nb214_jxH1(%esp),%xmm3 subps nb214_jyH1(%esp),%xmm4 subps nb214_jzH1(%esp),%xmm5 movaps %xmm0,nb214_dxOO(%esp) movaps %xmm1,nb214_dyOO(%esp) movaps %xmm2,nb214_dzOO(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb214_dxH1H1(%esp) movaps %xmm4,nb214_dyH1H1(%esp) movaps %xmm5,nb214_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,nb214_rsqOO(%esp) movaps %xmm3,nb214_rsqH1H1(%esp) movaps nb214_ixH1(%esp),%xmm0 movaps nb214_iyH1(%esp),%xmm1 movaps nb214_izH1(%esp),%xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 subps nb214_jxH2(%esp),%xmm0 subps nb214_jyH2(%esp),%xmm1 subps nb214_jzH2(%esp),%xmm2 subps nb214_jxM(%esp),%xmm3 subps nb214_jyM(%esp),%xmm4 subps nb214_jzM(%esp),%xmm5 movaps %xmm0,nb214_dxH1H2(%esp) movaps %xmm1,nb214_dyH1H2(%esp) movaps %xmm2,nb214_dzH1H2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb214_dxH1M(%esp) movaps %xmm4,nb214_dyH1M(%esp) movaps %xmm5,nb214_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,nb214_rsqH1H2(%esp) movaps %xmm3,nb214_rsqH1M(%esp) movaps nb214_ixH2(%esp),%xmm0 movaps nb214_iyH2(%esp),%xmm1 movaps nb214_izH2(%esp),%xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 subps nb214_jxH1(%esp),%xmm0 subps nb214_jyH1(%esp),%xmm1 subps nb214_jzH1(%esp),%xmm2 subps nb214_jxH2(%esp),%xmm3 subps nb214_jyH2(%esp),%xmm4 subps nb214_jzH2(%esp),%xmm5 movaps %xmm0,nb214_dxH2H1(%esp) movaps %xmm1,nb214_dyH2H1(%esp) movaps %xmm2,nb214_dzH2H1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb214_dxH2H2(%esp) movaps %xmm4,nb214_dyH2H2(%esp) movaps %xmm5,nb214_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,nb214_rsqH2H1(%esp) movaps %xmm3,nb214_rsqH2H2(%esp) movaps nb214_ixH2(%esp),%xmm0 movaps nb214_iyH2(%esp),%xmm1 movaps nb214_izH2(%esp),%xmm2 movaps nb214_ixM(%esp),%xmm3 movaps nb214_iyM(%esp),%xmm4 movaps nb214_izM(%esp),%xmm5 subps nb214_jxM(%esp),%xmm0 subps nb214_jyM(%esp),%xmm1 subps nb214_jzM(%esp),%xmm2 subps nb214_jxH1(%esp),%xmm3 subps nb214_jyH1(%esp),%xmm4 subps nb214_jzH1(%esp),%xmm5 movaps %xmm0,nb214_dxH2M(%esp) movaps %xmm1,nb214_dyH2M(%esp) movaps %xmm2,nb214_dzH2M(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb214_dxMH1(%esp) movaps %xmm4,nb214_dyMH1(%esp) movaps %xmm5,nb214_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,nb214_rsqH2M(%esp) movaps %xmm4,nb214_rsqMH1(%esp) movaps nb214_ixM(%esp),%xmm0 movaps nb214_iyM(%esp),%xmm1 movaps nb214_izM(%esp),%xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 subps nb214_jxH2(%esp),%xmm0 subps nb214_jyH2(%esp),%xmm1 subps nb214_jzH2(%esp),%xmm2 subps nb214_jxM(%esp),%xmm3 subps nb214_jyM(%esp),%xmm4 subps nb214_jzM(%esp),%xmm5 movaps %xmm0,nb214_dxMH2(%esp) movaps %xmm1,nb214_dyMH2(%esp) movaps %xmm2,nb214_dzMH2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb214_dxMM(%esp) movaps %xmm4,nb214_dyMM(%esp) movaps %xmm5,nb214_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,nb214_rsqMH2(%esp) movaps %xmm4,nb214_rsqMM(%esp) ## start by doing reciprocal for OO movaps nb214_rsqOO(%esp),%xmm7 rcpps %xmm7,%xmm2 movaps nb214_two(%esp),%xmm1 mulps %xmm2,%xmm7 subps %xmm7,%xmm1 mulps %xmm1,%xmm2 ## rinvsq movaps %xmm2,nb214_rinvsqOO(%esp) ## next step is invsqrt - do two at a time. rsqrtps nb214_rsqH1H1(%esp),%xmm1 rsqrtps nb214_rsqH1H2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb214_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb214_rsqH1H1(%esp),%xmm1 mulps nb214_rsqH1H2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb214_half(%esp),%xmm3 ## rinvH1H1 mulps nb214_half(%esp),%xmm7 ## rinvH1H2 movaps %xmm3,nb214_rinvH1H1(%esp) movaps %xmm7,nb214_rinvH1H2(%esp) rsqrtps nb214_rsqH1M(%esp),%xmm1 rsqrtps nb214_rsqH2H1(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb214_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb214_rsqH1M(%esp),%xmm1 mulps nb214_rsqH2H1(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb214_half(%esp),%xmm3 mulps nb214_half(%esp),%xmm7 movaps %xmm3,nb214_rinvH1M(%esp) movaps %xmm7,nb214_rinvH2H1(%esp) rsqrtps nb214_rsqH2H2(%esp),%xmm1 rsqrtps nb214_rsqH2M(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb214_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb214_rsqH2H2(%esp),%xmm1 mulps nb214_rsqH2M(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb214_half(%esp),%xmm3 mulps nb214_half(%esp),%xmm7 movaps %xmm3,nb214_rinvH2H2(%esp) movaps %xmm7,nb214_rinvH2M(%esp) rsqrtps nb214_rsqMH1(%esp),%xmm1 rsqrtps nb214_rsqMH2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb214_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb214_rsqMH1(%esp),%xmm1 mulps nb214_rsqMH2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb214_half(%esp),%xmm3 mulps nb214_half(%esp),%xmm7 movaps %xmm3,nb214_rinvMH1(%esp) movaps %xmm7,nb214_rinvMH2(%esp) rsqrtps nb214_rsqMM(%esp),%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb214_three(%esp),%xmm3 mulps nb214_rsqMM(%esp),%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb214_half(%esp),%xmm3 movaps %xmm3,nb214_rinvMM(%esp) ## start with OO LJ interaction movaps nb214_rinvsqOO(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm1,%xmm1 ## rinv4 mulps %xmm0,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinvtwelve mulps nb214_c6(%esp),%xmm1 mulps nb214_c12(%esp),%xmm2 movaps %xmm2,%xmm4 subps %xmm1,%xmm4 addps nb214_Vvdwtot(%esp),%xmm4 mulps nb214_six(%esp),%xmm1 mulps nb214_twelve(%esp),%xmm2 movaps %xmm4,nb214_Vvdwtot(%esp) subps %xmm1,%xmm2 mulps %xmm2,%xmm0 ## fscal movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb214_dxOO(%esp),%xmm0 mulps nb214_dyOO(%esp),%xmm1 mulps nb214_dzOO(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb214_fixO(%esp),%xmm0 addps nb214_fiyO(%esp),%xmm1 addps nb214_fizO(%esp),%xmm2 movaps %xmm3,nb214_fjxO(%esp) movaps %xmm4,nb214_fjyO(%esp) movaps %xmm5,nb214_fjzO(%esp) movaps %xmm0,nb214_fixO(%esp) movaps %xmm1,nb214_fiyO(%esp) movaps %xmm2,nb214_fizO(%esp) ## Coulomb interactions ## start with H1-H1 interaction movaps nb214_rinvH1H1(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb214_krf(%esp),%xmm5 mulps %xmm0,%xmm0 ## xmm0=rinvsq mulps nb214_rsqH1H1(%esp),%xmm5 ## xmm5=krsq
⌨️ 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?