nb_kernel302_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,970 行 · 第 1/5 页
S
1,970 行
unpckhps %xmm5,%xmm2 ## xmm2= jxOc jxOd jyOc jyOd unpckhps %xmm6,%xmm3 ## xmm3= jxH1c jxH1d jyH1c jyH1d movaps %xmm4,%xmm5 movaps %xmm0,%xmm6 unpcklps %xmm7,%xmm4 ## xmm4= jxH2a jxH2b jyH2a jyH2b unpckhps %xmm7,%xmm5 ## xmm5= jxH2c jxH2d jyH2c jyH2d movaps %xmm1,%xmm7 movlhps %xmm2,%xmm0 ## xmm0= jxOa jxOb jxOc jxOd movaps %xmm0,nb302_jxO(%esp) movhlps %xmm6,%xmm2 ## xmm2= jyOa jyOb jyOc jyOd movaps %xmm2,nb302_jyO(%esp) movlhps %xmm3,%xmm1 movaps %xmm1,nb302_jxH1(%esp) movhlps %xmm7,%xmm3 movaps %xmm4,%xmm6 movaps %xmm3,nb302_jyH1(%esp) movlhps %xmm5,%xmm4 movaps %xmm4,nb302_jxH2(%esp) movhlps %xmm6,%xmm5 movaps %xmm5,nb302_jyH2(%esp) movss 8(%esi,%eax,4),%xmm0 movss 20(%esi,%eax,4),%xmm1 movss 32(%esi,%eax,4),%xmm2 movss 8(%esi,%ecx,4),%xmm3 movss 20(%esi,%ecx,4),%xmm4 movss 32(%esi,%ecx,4),%xmm5 movhps 4(%esi,%ebx,4),%xmm0 movhps 16(%esi,%ebx,4),%xmm1 movhps 28(%esi,%ebx,4),%xmm2 movhps 4(%esi,%edx,4),%xmm3 movhps 16(%esi,%edx,4),%xmm4 movhps 28(%esi,%edx,4),%xmm5 shufps $204,%xmm3,%xmm0 ## constant 11001100 shufps $204,%xmm4,%xmm1 ## constant 11001100 shufps $204,%xmm5,%xmm2 ## constant 11001100 movaps %xmm0,nb302_jzO(%esp) movaps %xmm1,nb302_jzH1(%esp) movaps %xmm2,nb302_jzH2(%esp) movaps nb302_ixO(%esp),%xmm0 movaps nb302_iyO(%esp),%xmm1 movaps nb302_izO(%esp),%xmm2 movaps nb302_ixO(%esp),%xmm3 movaps nb302_iyO(%esp),%xmm4 movaps nb302_izO(%esp),%xmm5 subps nb302_jxO(%esp),%xmm0 subps nb302_jyO(%esp),%xmm1 subps nb302_jzO(%esp),%xmm2 subps nb302_jxH1(%esp),%xmm3 subps nb302_jyH1(%esp),%xmm4 subps nb302_jzH1(%esp),%xmm5 movaps %xmm0,nb302_dxOO(%esp) movaps %xmm1,nb302_dyOO(%esp) movaps %xmm2,nb302_dzOO(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb302_dxOH1(%esp) movaps %xmm4,nb302_dyOH1(%esp) movaps %xmm5,nb302_dzOH1(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb302_rsqOO(%esp) movaps %xmm3,nb302_rsqOH1(%esp) movaps nb302_ixO(%esp),%xmm0 movaps nb302_iyO(%esp),%xmm1 movaps nb302_izO(%esp),%xmm2 movaps nb302_ixH1(%esp),%xmm3 movaps nb302_iyH1(%esp),%xmm4 movaps nb302_izH1(%esp),%xmm5 subps nb302_jxH2(%esp),%xmm0 subps nb302_jyH2(%esp),%xmm1 subps nb302_jzH2(%esp),%xmm2 subps nb302_jxO(%esp),%xmm3 subps nb302_jyO(%esp),%xmm4 subps nb302_jzO(%esp),%xmm5 movaps %xmm0,nb302_dxOH2(%esp) movaps %xmm1,nb302_dyOH2(%esp) movaps %xmm2,nb302_dzOH2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb302_dxH1O(%esp) movaps %xmm4,nb302_dyH1O(%esp) movaps %xmm5,nb302_dzH1O(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb302_rsqOH2(%esp) movaps %xmm3,nb302_rsqH1O(%esp) movaps nb302_ixH1(%esp),%xmm0 movaps nb302_iyH1(%esp),%xmm1 movaps nb302_izH1(%esp),%xmm2 movaps nb302_ixH1(%esp),%xmm3 movaps nb302_iyH1(%esp),%xmm4 movaps nb302_izH1(%esp),%xmm5 subps nb302_jxH1(%esp),%xmm0 subps nb302_jyH1(%esp),%xmm1 subps nb302_jzH1(%esp),%xmm2 subps nb302_jxH2(%esp),%xmm3 subps nb302_jyH2(%esp),%xmm4 subps nb302_jzH2(%esp),%xmm5 movaps %xmm0,nb302_dxH1H1(%esp) movaps %xmm1,nb302_dyH1H1(%esp) movaps %xmm2,nb302_dzH1H1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb302_dxH1H2(%esp) movaps %xmm4,nb302_dyH1H2(%esp) movaps %xmm5,nb302_dzH1H2(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb302_rsqH1H1(%esp) movaps %xmm3,nb302_rsqH1H2(%esp) movaps nb302_ixH2(%esp),%xmm0 movaps nb302_iyH2(%esp),%xmm1 movaps nb302_izH2(%esp),%xmm2 movaps nb302_ixH2(%esp),%xmm3 movaps nb302_iyH2(%esp),%xmm4 movaps nb302_izH2(%esp),%xmm5 subps nb302_jxO(%esp),%xmm0 subps nb302_jyO(%esp),%xmm1 subps nb302_jzO(%esp),%xmm2 subps nb302_jxH1(%esp),%xmm3 subps nb302_jyH1(%esp),%xmm4 subps nb302_jzH1(%esp),%xmm5 movaps %xmm0,nb302_dxH2O(%esp) movaps %xmm1,nb302_dyH2O(%esp) movaps %xmm2,nb302_dzH2O(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb302_dxH2H1(%esp) movaps %xmm4,nb302_dyH2H1(%esp) movaps %xmm5,nb302_dzH2H1(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm3,%xmm4 addps %xmm5,%xmm4 movaps %xmm0,nb302_rsqH2O(%esp) movaps %xmm4,nb302_rsqH2H1(%esp) movaps nb302_ixH2(%esp),%xmm0 movaps nb302_iyH2(%esp),%xmm1 movaps nb302_izH2(%esp),%xmm2 subps nb302_jxH2(%esp),%xmm0 subps nb302_jyH2(%esp),%xmm1 subps nb302_jzH2(%esp),%xmm2 movaps %xmm0,nb302_dxH2H2(%esp) movaps %xmm1,nb302_dyH2H2(%esp) movaps %xmm2,nb302_dzH2H2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 movaps %xmm0,nb302_rsqH2H2(%esp) ## start doing invsqrt use rsq values in xmm0, xmm4 rsqrtps %xmm0,%xmm1 rsqrtps %xmm4,%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb302_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps %xmm0,%xmm1 mulps %xmm4,%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb302_half(%esp),%xmm3 ## rinvH2H2 mulps nb302_half(%esp),%xmm7 ## rinvH2H1 movaps %xmm3,nb302_rinvH2H2(%esp) movaps %xmm7,nb302_rinvH2H1(%esp) rsqrtps nb302_rsqOO(%esp),%xmm1 rsqrtps nb302_rsqOH1(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb302_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb302_rsqOO(%esp),%xmm1 mulps nb302_rsqOH1(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb302_half(%esp),%xmm3 mulps nb302_half(%esp),%xmm7 movaps %xmm3,nb302_rinvOO(%esp) movaps %xmm7,nb302_rinvOH1(%esp) rsqrtps nb302_rsqOH2(%esp),%xmm1 rsqrtps nb302_rsqH1O(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb302_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb302_rsqOH2(%esp),%xmm1 mulps nb302_rsqH1O(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb302_half(%esp),%xmm3 mulps nb302_half(%esp),%xmm7 movaps %xmm3,nb302_rinvOH2(%esp) movaps %xmm7,nb302_rinvH1O(%esp) rsqrtps nb302_rsqH1H1(%esp),%xmm1 rsqrtps nb302_rsqH1H2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb302_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb302_rsqH1H1(%esp),%xmm1 mulps nb302_rsqH1H2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb302_half(%esp),%xmm3 mulps nb302_half(%esp),%xmm7 movaps %xmm3,nb302_rinvH1H1(%esp) movaps %xmm7,nb302_rinvH1H2(%esp) rsqrtps nb302_rsqH2O(%esp),%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb302_three(%esp),%xmm3 mulps nb302_rsqH2O(%esp),%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb302_half(%esp),%xmm3 movaps %xmm3,nb302_rinvH2O(%esp) ## start with OO interaction movaps nb302_rinvOO(%esp),%xmm0 movaps %xmm0,%xmm1 mulps nb302_rsqOO(%esp),%xmm1 ## xmm1=r mulps nb302_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 nb302_VFtab(%ebp),%esi 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 nb302_two(%esp),%xmm7 ## two*Heps2 movaps nb302_qqOO(%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 ## increment vcoul - then we can get rid of mm5 ## update vctot addps nb302_vctot(%esp),%xmm5 xorps %xmm2,%xmm2 movaps %xmm5,nb302_vctot(%esp) mulps nb302_tsc(%esp),%xmm3 subps %xmm3,%xmm2 mulps %xmm2,%xmm0 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb302_dxOO(%esp),%xmm0 mulps nb302_dyOO(%esp),%xmm1 mulps nb302_dzOO(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb302_fixO(%esp),%xmm0 addps nb302_fiyO(%esp),%xmm1 addps nb302_fizO(%esp),%xmm2 movaps %xmm3,nb302_fjxO(%esp) movaps %xmm4,nb302_fjyO(%esp) movaps %xmm5,nb302_fjzO(%esp) movaps %xmm0,nb302_fixO(%esp) movaps %xmm1,nb302_fiyO(%esp) movaps %xmm2,nb302_fizO(%esp) ## O-H1 interaction movaps nb302_rinvOH1(%esp),%xmm0 movaps %xmm0,%xmm1 mulps nb302_rsqOH1(%esp),%xmm1 ## xmm1=r mulps nb302_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
⌨️ 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?