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
movhlps %xmm6,%xmm2 movaps %xmm2,nb304_jyH1(%esp) movlhps %xmm3,%xmm1 movaps %xmm1,nb304_jxH2(%esp) movhlps %xmm7,%xmm3 movaps %xmm4,%xmm6 movaps %xmm3,nb304_jyH2(%esp) movlhps %xmm5,%xmm4 movaps %xmm4,nb304_jxM(%esp) movhlps %xmm6,%xmm5 movaps %xmm5,nb304_jyM(%esp) movss 20(%esi,%eax,4),%xmm0 movss 32(%esi,%eax,4),%xmm1 movss 44(%esi,%eax,4),%xmm2 movss 20(%esi,%ecx,4),%xmm3 movss 32(%esi,%ecx,4),%xmm4 movss 44(%esi,%ecx,4),%xmm5 movhps 16(%esi,%ebx,4),%xmm0 movhps 28(%esi,%ebx,4),%xmm1 movhps 40(%esi,%ebx,4),%xmm2 movhps 16(%esi,%edx,4),%xmm3 movhps 28(%esi,%edx,4),%xmm4 movhps 40(%esi,%edx,4),%xmm5 shufps $204,%xmm3,%xmm0 ## constant 11001100 shufps $204,%xmm4,%xmm1 ## constant 11001100 shufps $204,%xmm5,%xmm2 ## constant 11001100 movaps %xmm0,nb304_jzH1(%esp) movaps %xmm1,nb304_jzH2(%esp) movaps %xmm2,nb304_jzM(%esp) movaps nb304_ixH1(%esp),%xmm0 movaps nb304_iyH1(%esp),%xmm1 movaps nb304_izH1(%esp),%xmm2 movaps nb304_ixH1(%esp),%xmm3 movaps nb304_iyH1(%esp),%xmm4 movaps nb304_izH1(%esp),%xmm5 subps nb304_jxH1(%esp),%xmm0 subps nb304_jyH1(%esp),%xmm1 subps nb304_jzH1(%esp),%xmm2 subps nb304_jxH2(%esp),%xmm3 subps nb304_jyH2(%esp),%xmm4 subps nb304_jzH2(%esp),%xmm5 movaps %xmm0,nb304_dxH1H1(%esp) movaps %xmm1,nb304_dyH1H1(%esp) movaps %xmm2,nb304_dzH1H1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb304_dxH1H2(%esp) movaps %xmm4,nb304_dyH1H2(%esp) movaps %xmm5,nb304_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,nb304_rsqH1H1(%esp) movaps %xmm3,nb304_rsqH1H2(%esp) movaps nb304_ixH1(%esp),%xmm0 movaps nb304_iyH1(%esp),%xmm1 movaps nb304_izH1(%esp),%xmm2 movaps nb304_ixH2(%esp),%xmm3 movaps nb304_iyH2(%esp),%xmm4 movaps nb304_izH2(%esp),%xmm5 subps nb304_jxM(%esp),%xmm0 subps nb304_jyM(%esp),%xmm1 subps nb304_jzM(%esp),%xmm2 subps nb304_jxH1(%esp),%xmm3 subps nb304_jyH1(%esp),%xmm4 subps nb304_jzH1(%esp),%xmm5 movaps %xmm0,nb304_dxH1M(%esp) movaps %xmm1,nb304_dyH1M(%esp) movaps %xmm2,nb304_dzH1M(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb304_dxH2H1(%esp) movaps %xmm4,nb304_dyH2H1(%esp) movaps %xmm5,nb304_dzH2H1(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb304_rsqH1M(%esp) movaps %xmm3,nb304_rsqH2H1(%esp) movaps nb304_ixH2(%esp),%xmm0 movaps nb304_iyH2(%esp),%xmm1 movaps nb304_izH2(%esp),%xmm2 movaps nb304_ixH2(%esp),%xmm3 movaps nb304_iyH2(%esp),%xmm4 movaps nb304_izH2(%esp),%xmm5 subps nb304_jxH2(%esp),%xmm0 subps nb304_jyH2(%esp),%xmm1 subps nb304_jzH2(%esp),%xmm2 subps nb304_jxM(%esp),%xmm3 subps nb304_jyM(%esp),%xmm4 subps nb304_jzM(%esp),%xmm5 movaps %xmm0,nb304_dxH2H2(%esp) movaps %xmm1,nb304_dyH2H2(%esp) movaps %xmm2,nb304_dzH2H2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb304_dxH2M(%esp) movaps %xmm4,nb304_dyH2M(%esp) movaps %xmm5,nb304_dzH2M(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb304_rsqH2H2(%esp) movaps %xmm3,nb304_rsqH2M(%esp) movaps nb304_ixM(%esp),%xmm0 movaps nb304_iyM(%esp),%xmm1 movaps nb304_izM(%esp),%xmm2 movaps nb304_ixM(%esp),%xmm3 movaps nb304_iyM(%esp),%xmm4 movaps nb304_izM(%esp),%xmm5 subps nb304_jxH1(%esp),%xmm0 subps nb304_jyH1(%esp),%xmm1 subps nb304_jzH1(%esp),%xmm2 subps nb304_jxH2(%esp),%xmm3 subps nb304_jyH2(%esp),%xmm4 subps nb304_jzH2(%esp),%xmm5 movaps %xmm0,nb304_dxMH1(%esp) movaps %xmm1,nb304_dyMH1(%esp) movaps %xmm2,nb304_dzMH1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb304_dxMH2(%esp) movaps %xmm4,nb304_dyMH2(%esp) movaps %xmm5,nb304_dzMH2(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm3,%xmm4 addps %xmm5,%xmm4 movaps %xmm0,nb304_rsqMH1(%esp) movaps %xmm4,nb304_rsqMH2(%esp) movaps nb304_ixM(%esp),%xmm0 movaps nb304_iyM(%esp),%xmm1 movaps nb304_izM(%esp),%xmm2 subps nb304_jxM(%esp),%xmm0 subps nb304_jyM(%esp),%xmm1 subps nb304_jzM(%esp),%xmm2 movaps %xmm0,nb304_dxMM(%esp) movaps %xmm1,nb304_dyMM(%esp) movaps %xmm2,nb304_dzMM(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 movaps %xmm0,nb304_rsqMM(%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 nb304_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 nb304_half(%esp),%xmm3 ## rinvMM mulps nb304_half(%esp),%xmm7 ## rinvMH2 movaps %xmm3,nb304_rinvMM(%esp) movaps %xmm7,nb304_rinvMH2(%esp) rsqrtps nb304_rsqH1H1(%esp),%xmm1 rsqrtps nb304_rsqH1H2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb304_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb304_rsqH1H1(%esp),%xmm1 mulps nb304_rsqH1H2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb304_half(%esp),%xmm3 mulps nb304_half(%esp),%xmm7 movaps %xmm3,nb304_rinvH1H1(%esp) movaps %xmm7,nb304_rinvH1H2(%esp) rsqrtps nb304_rsqH1M(%esp),%xmm1 rsqrtps nb304_rsqH2H1(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb304_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb304_rsqH1M(%esp),%xmm1 mulps nb304_rsqH2H1(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb304_half(%esp),%xmm3 mulps nb304_half(%esp),%xmm7 movaps %xmm3,nb304_rinvH1M(%esp) movaps %xmm7,nb304_rinvH2H1(%esp) rsqrtps nb304_rsqH2H2(%esp),%xmm1 rsqrtps nb304_rsqH2M(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb304_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb304_rsqH2H2(%esp),%xmm1 mulps nb304_rsqH2M(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb304_half(%esp),%xmm3 mulps nb304_half(%esp),%xmm7 movaps %xmm3,nb304_rinvH2H2(%esp) movaps %xmm7,nb304_rinvH2M(%esp) rsqrtps nb304_rsqMH1(%esp),%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb304_three(%esp),%xmm3 mulps nb304_rsqMH1(%esp),%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb304_half(%esp),%xmm3 movaps %xmm3,nb304_rinvMH1(%esp) ## start with H1-H1 interaction movaps nb304_rinvH1H1(%esp),%xmm0 movaps %xmm0,%xmm1 mulps nb304_rsqH1H1(%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 %eax,%mm0 movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movl nb304_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 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 ## increment vcoul - then we can get rid of mm5 ## update vctot addps nb304_vctot(%esp),%xmm5 xorps %xmm2,%xmm2 movaps %xmm5,nb304_vctot(%esp) mulps nb304_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 nb304_dxH1H1(%esp),%xmm0 mulps nb304_dyH1H1(%esp),%xmm1 mulps nb304_dzH1H1(%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_fjxH1(%esp) movaps %xmm4,nb304_fjyH1(%esp) movaps %xmm5,nb304_fjzH1(%esp) movaps %xmm0,nb304_fixH1(%esp) movaps %xmm1,nb304_fiyH1(%esp) movaps %xmm2,nb304_fizH1(%esp) ## H1-H2 interaction movaps nb304_rinvH1H2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps nb304_rsqH1H2(%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
⌨️ 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?