nb_kernel333_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,082 行 · 第 1/5 页
S
2,082 行
shufps $0,%xmm0,%xmm2 shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## constant 10001000 shufps $136,%xmm5,%xmm0 ## constant 10001000 shufps $221,%xmm5,%xmm1 ## constant 11011101 ## move ixO-izO to xmm4-xmm6 movaps nb333_ixO(%esp),%xmm4 movaps nb333_iyO(%esp),%xmm5 movaps nb333_izO(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb333_dxO(%esp) movaps %xmm5,nb333_dyO(%esp) movaps %xmm6,nb333_dzO(%esp) ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 movaps %xmm4,%xmm7 ## rsqO in xmm7 ## move ixH1-izH1 to xmm4-xmm6 movaps nb333_ixH1(%esp),%xmm4 movaps nb333_iyH1(%esp),%xmm5 movaps nb333_izH1(%esp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## store dr movaps %xmm4,nb333_dxH1(%esp) movaps %xmm5,nb333_dyH1(%esp) movaps %xmm6,nb333_dzH1(%esp) ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm6 addps %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixH2-izH2 to xmm3-xmm5 movaps nb333_ixH2(%esp),%xmm3 movaps nb333_iyH2(%esp),%xmm4 movaps nb333_izH2(%esp),%xmm5 ## calc dr subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 ## store dr movaps %xmm3,nb333_dxH2(%esp) movaps %xmm4,nb333_dyH2(%esp) movaps %xmm5,nb333_dzH2(%esp) ## square it mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm4,%xmm5 addps %xmm3,%xmm5 ## move ixM-izM to xmm2-xmm4 movaps nb333_iyM(%esp),%xmm3 movaps nb333_izM(%esp),%xmm4 subps %xmm1,%xmm3 subps %xmm2,%xmm4 movaps nb333_ixM(%esp),%xmm2 subps %xmm0,%xmm2 ## store dr movaps %xmm2,nb333_dxM(%esp) movaps %xmm3,nb333_dyM(%esp) movaps %xmm4,nb333_dzM(%esp) ## square it mulps %xmm2,%xmm2 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 addps %xmm3,%xmm4 addps %xmm2,%xmm4 ## rsqM in xmm4, rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 ## rsqH1 - seed in xmm2 rsqrtps %xmm6,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb333_three(%esp),%xmm0 mulps %xmm6,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb333_half(%esp),%xmm0 movaps %xmm0,nb333_rinvH1(%esp) ## rinvH1 mulps %xmm0,%xmm6 movaps %xmm6,nb333_rH1(%esp) ## rsqH2 - seed to xmm2 rsqrtps %xmm5,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb333_three(%esp),%xmm0 mulps %xmm5,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb333_half(%esp),%xmm0 movaps %xmm0,nb333_rinvH2(%esp) ## rinvH2 mulps %xmm0,%xmm5 movaps %xmm5,nb333_rH2(%esp) ## rsqM - seed to xmm2 rsqrtps %xmm4,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb333_three(%esp),%xmm0 mulps %xmm4,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb333_half(%esp),%xmm0 movaps %xmm0,nb333_rinvM(%esp) ## rinvM mulps %xmm0,%xmm4 movaps %xmm4,nb333_rM(%esp) ## Do the O LJ table interaction directly. rsqrtps %xmm7,%xmm2 movaps %xmm2,%xmm3 mulps %xmm2,%xmm2 movaps nb333_three(%esp),%xmm0 mulps %xmm7,%xmm2 ## rsq*lu*lu subps %xmm2,%xmm0 ## constant 30-rsq*lu*lu mulps %xmm3,%xmm0 ## lu*(3-rsq*lu*lu) mulps nb333_half(%esp),%xmm0 ## rinv movaps %xmm0,%xmm1 mulps %xmm7,%xmm1 ## xmm1=r mulps nb333_tsc(%esp),%xmm1 ## r*tabscale 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 nb333_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 coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 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 coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## dispersion table YFGH ready in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp mulps nb333_two(%esp),%xmm7 ## two*Heps2 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV movaps nb333_c6(%esp),%xmm4 mulps %xmm4,%xmm7 ## fijD mulps %xmm4,%xmm5 ## Vvdw6 ## put scalar force on stack (borrow rinvO) ## Update Vvdwtot directly addps nb333_Vvdwtot(%esp),%xmm5 movaps %xmm7,nb333_rinvO(%esp) ## fscal movaps %xmm5,nb333_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 movhps 32(%esi,%edx,4),%xmm7 ## got half coulomb table movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101 movlps 40(%esi,%eax,4),%xmm7 movlps 40(%esi,%ecx,4),%xmm3 movhps 40(%esi,%ebx,4),%xmm7 movhps 40(%esi,%edx,4),%xmm3 ## other half of coulomb table movaps %xmm7,%xmm6 shufps $136,%xmm3,%xmm6 ## constant 10001000 shufps $221,%xmm3,%xmm7 ## constant 11011101 ## repulsion table YFGH ready in xmm4-xmm7 mulps %xmm1,%xmm6 ## xmm6=Geps mulps %xmm2,%xmm7 ## xmm7=Heps2 addps %xmm6,%xmm5 addps %xmm7,%xmm5 ## xmm5=Fp mulps nb333_two(%esp),%xmm7 ## two*Heps2 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV movaps nb333_c12(%esp),%xmm4 mulps %xmm4,%xmm7 ## fijR mulps %xmm4,%xmm5 ## Vvdw12 addps nb333_rinvO(%esp),%xmm7 ## fscal was temp. stored in rinvO addps nb333_Vvdwtot(%esp),%xmm5 movaps %xmm5,nb333_Vvdwtot(%esp) xorps %xmm1,%xmm1 mulps nb333_tsc(%esp),%xmm7 mulps %xmm0,%xmm7 subps %xmm7,%xmm1 ## fscal movaps nb333_dxO(%esp),%xmm3 movaps nb333_dyO(%esp),%xmm4 movaps nb333_dzO(%esp),%xmm5 mulps %xmm1,%xmm3 mulps %xmm1,%xmm4 mulps %xmm1,%xmm5 ## tx in xmm3-xmm5 ## update O forces movaps nb333_fixO(%esp),%xmm0 movaps nb333_fiyO(%esp),%xmm1 movaps nb333_fizO(%esp),%xmm2 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 movaps %xmm0,nb333_fixO(%esp) movaps %xmm1,nb333_fiyO(%esp) movaps %xmm2,nb333_fizO(%esp) ## update j forces with water O movaps %xmm3,nb333_fjx(%esp) movaps %xmm4,nb333_fjy(%esp) movaps %xmm5,nb333_fjz(%esp) ## Do H1 interaction movl nb333_VFtab(%ebp),%esi movaps nb333_rH1(%esp),%xmm7 mulps nb333_tsc(%esp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%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 ## 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 nb333_two(%esp),%xmm7 ## two*Heps2 movaps nb333_qqH(%esp),%xmm0 addps %xmm6,%xmm7 addps %xmm5,%xmm7 ## xmm7=FF mulps %xmm1,%xmm5 ## xmm5=eps*Fp addps %xmm4,%xmm5 ## xmm5=VV mulps %xmm0,%xmm5 ## vcoul=qq*VV mulps %xmm0,%xmm7 ## fijC=FF*qq ## at this point mm5 contains vcoul and xmm7 fijC ## increment vcoul xorps %xmm4,%xmm4 addps nb333_vctot(%esp),%xmm5 mulps nb333_rinvH1(%esp),%xmm7 movaps %xmm5,nb333_vctot(%esp) mulps nb333_tsc(%esp),%xmm7 subps %xmm7,%xmm4 movaps nb333_dxH1(%esp),%xmm0 movaps nb333_dyH1(%esp),%xmm1 movaps nb333_dzH1(%esp),%xmm2 mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## update H1 forces movaps nb333_fixH1(%esp),%xmm3 movaps nb333_fiyH1(%esp),%xmm4 movaps nb333_fizH1(%esp),%xmm7 addps %xmm0,%xmm3 addps %xmm1,%xmm4 addps %xmm2,%xmm7 movaps %xmm3,nb333_fixH1(%esp) movaps %xmm4,nb333_fiyH1(%esp) movaps %xmm7,nb333_fizH1(%esp) ## update j forces with water H1 addps nb333_fjx(%esp),%xmm0 addps nb333_fjy(%esp),%xmm1 addps nb333_fjz(%esp),%xmm2 movaps %xmm0,nb333_fjx(%esp) movaps %xmm1,nb333_fjy(%esp) movaps %xmm2,nb333_fjz(%esp) ## Done with H1, do H2 interactions movaps nb333_rH2(%esp),%xmm7 mulps nb333_tsc(%esp),%xmm7 movhlps %xmm7,%xmm4 cvttps2pi %xmm7,%mm6 cvttps2pi %xmm4,%mm7 ## mm6/mm7 contain lu indices cvtpi2ps %mm6,%xmm3 cvtpi2ps %mm7,%xmm4 movlhps %xmm4,%xmm3 subps %xmm3,%xmm7 movaps %xmm7,%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 ## constant 10001000 shufps $221,%xmm7,%xmm5 ## constant 11011101⌨️ 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?