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
subps %xmm3,%xmm1 movaps %xmm1,%xmm0 movaps %xmm1,%xmm2 movaps nb334_fjxH1(%esp),%xmm3 movaps nb334_fjyH1(%esp),%xmm4 movaps nb334_fjzH1(%esp),%xmm5 mulps nb334_dxH2H1(%esp),%xmm0 mulps nb334_dyH2H1(%esp),%xmm1 mulps nb334_dzH2H1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb334_fixH2(%esp),%xmm0 addps nb334_fiyH2(%esp),%xmm1 addps nb334_fizH2(%esp),%xmm2 movaps %xmm3,nb334_fjxH1(%esp) movaps %xmm4,nb334_fjyH1(%esp) movaps %xmm5,nb334_fjzH1(%esp) movaps %xmm0,nb334_fixH2(%esp) movaps %xmm1,nb334_fiyH2(%esp) movaps %xmm2,nb334_fizH2(%esp) ## H2-H2 interaction movaps nb334_rinvH2H2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps nb334_rsqH2H2(%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 %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 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 ## 10001000 shufps $221,%xmm7,%xmm5 ## 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 ## 10001000 shufps $221,%xmm3,%xmm7 ## 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 nb334_two(%esp),%xmm7 ## two*Heps2 movaps nb334_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 addps nb334_vctot(%esp),%xmm5 movaps %xmm5,nb334_vctot(%esp) xorps %xmm1,%xmm1 mulps nb334_tsc(%esp),%xmm3 mulps %xmm0,%xmm3 subps %xmm3,%xmm1 movaps %xmm1,%xmm0 movaps %xmm1,%xmm2 movaps nb334_fjxH2(%esp),%xmm3 movaps nb334_fjyH2(%esp),%xmm4 movaps nb334_fjzH2(%esp),%xmm5 mulps nb334_dxH2H2(%esp),%xmm0 mulps nb334_dyH2H2(%esp),%xmm1 mulps nb334_dzH2H2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb334_fixH2(%esp),%xmm0 addps nb334_fiyH2(%esp),%xmm1 addps nb334_fizH2(%esp),%xmm2 movaps %xmm3,nb334_fjxH2(%esp) movaps %xmm4,nb334_fjyH2(%esp) movaps %xmm5,nb334_fjzH2(%esp) movaps %xmm0,nb334_fixH2(%esp) movaps %xmm1,nb334_fiyH2(%esp) movaps %xmm2,nb334_fizH2(%esp) ## H2-M interaction movaps nb334_rinvH2M(%esp),%xmm0 movaps %xmm0,%xmm1 mulps nb334_rsqH2M(%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 %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 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 ## 10001000 shufps $221,%xmm7,%xmm5 ## 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 ## 10001000 shufps $221,%xmm3,%xmm7 ## 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 nb334_two(%esp),%xmm7 ## two*Heps2 movaps nb334_qqMH(%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 addps nb334_vctot(%esp),%xmm5 movaps %xmm5,nb334_vctot(%esp) xorps %xmm1,%xmm1 mulps nb334_tsc(%esp),%xmm3 mulps %xmm0,%xmm3 subps %xmm3,%xmm1 movaps %xmm1,%xmm0 movaps %xmm1,%xmm2 movaps nb334_fjxM(%esp),%xmm3 movaps nb334_fjyM(%esp),%xmm4 movaps nb334_fjzM(%esp),%xmm5 mulps nb334_dxH2M(%esp),%xmm0 mulps nb334_dyH2M(%esp),%xmm1 mulps nb334_dzH2M(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb334_fixH2(%esp),%xmm0 addps nb334_fiyH2(%esp),%xmm1 addps nb334_fizH2(%esp),%xmm2 movaps %xmm3,nb334_fjxM(%esp) movaps %xmm4,nb334_fjyM(%esp) movaps %xmm5,nb334_fjzM(%esp) movaps %xmm0,nb334_fixH2(%esp) movaps %xmm1,nb334_fiyH2(%esp) movaps %xmm2,nb334_fizH2(%esp) ## M-H1 interaction movaps nb334_rinvMH1(%esp),%xmm0 movaps %xmm0,%xmm1 mulps nb334_rsqMH1(%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 %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 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 ## 10001000 shufps $221,%xmm7,%xmm5 ## 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 ## 10001000 shufps $221,%xmm3,%xmm7 ## 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 nb334_two(%esp),%xmm7 ## two*Heps2 movaps nb334_qqMH(%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 addps nb334_vctot(%esp),%xmm5 movaps %xmm5,nb334_vctot(%esp) xorps %xmm1,%xmm1 mulps nb334_tsc(%esp),%xmm3 mulps %xmm0,%xmm3 subps %xmm3,%xmm1 movaps %xmm1,%xmm0 movaps %xmm1,%xmm2 movaps nb334_fjxH1(%esp),%xmm3 movaps nb334_fjyH1(%esp),%xmm4 movaps nb334_fjzH1(%esp),%xmm5 mulps nb334_dxMH1(%esp),%xmm0 mulps nb334_dyMH1(%esp),%xmm1 mulps nb334_dzMH1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb334_fixM(%esp),%xmm0 addps nb334_fiyM(%esp),%xmm1 addps nb334_fizM(%esp),%xmm2 movaps %xmm3,nb334_fjxH1(%esp) movaps %xmm4,nb334_fjyH1(%esp) movaps %xmm5,nb334_fjzH1(%esp) movaps %xmm0,nb334_fixM(%esp) movaps %xmm1,nb334_fiyM(%esp) movaps %xmm2,nb334_fizM(%esp) ## M-H2 interaction movaps nb334_rinvMH2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps nb334_rsqMH2(%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 %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 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 ## 10001000 shufps $221,%xmm7,%xmm5 ## 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 ## 10001000 shufps $221,%xmm3,%xmm7 ## 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 nb334_two(%esp),%xmm7 ## two*Heps2 movaps nb334_qqMH(%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 addps nb334_vctot(%esp),%xmm5 movaps %xmm5,nb334_vctot(%esp) xorps %xmm1,%xmm1 mulps nb334_tsc(%esp),%xmm3 mulps %xmm0,%xmm3 subps %xmm3,%xmm1 movaps %xmm1,%xmm0 movaps %xmm1,%xmm2 movaps nb334_fjxH2(%esp),%xmm3 movaps nb334_fjyH2(%esp),%xmm4 movaps nb334_fjzH2(%esp),%xmm5 mulps nb334_dxMH2(%esp),%xmm0 mulps nb334_dyMH2(%esp),%xmm1 mulps nb334_dzMH2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb334_fixM(%esp),%xmm0 addps nb334_fiyM(%esp),%xmm1 addps nb334_fizM(%esp),%xmm2 movaps %xmm3,nb334_fjxH2(%esp) movaps %xmm4,nb334_fjyH2(%esp) movaps %xmm5,nb334_fjzH2(%esp) movaps %xmm0,nb334_fixM(%esp) movaps %xmm1,nb334_fiyM(%esp) movaps %xmm2,nb334_fizM(%esp) ## M-M interaction movaps nb334_rinvMM(%esp),%xmm0 movaps %xmm0,%xmm1 mulps nb334_rsqMM(%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
⌨️ 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?