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
xorps %xmm5,%xmm5 movss (%esi,%eax,4),%xmm3 ## jxO - - - movss 4(%esi,%eax,4),%xmm4 ## jyO - - - movss 8(%esi,%eax,4),%xmm5 ## jzO - - - movlps 12(%esi,%eax,4),%xmm6 ## xmm6 = jxH1 jyH1 - - movss 20(%esi,%eax,4),%xmm7 ## xmm7 = jzH1 - - - movhps 24(%esi,%eax,4),%xmm6 ## xmm6 = jxH1 jyH1 jxH2 jyH2 movss 32(%esi,%eax,4),%xmm2 ## xmm2 = jzH2 - - - ## have all coords, time for some shuffling. shufps $216,%xmm6,%xmm6 ## constant 11011000 ;# xmm6 = jxH1 jxH2 jyH1 jyH2 unpcklps %xmm2,%xmm7 ## xmm7 = jzH1 jzH2 - - movaps nb302_ixO(%esp),%xmm0 movaps nb302_iyO(%esp),%xmm1 movaps nb302_izO(%esp),%xmm2 movlhps %xmm6,%xmm3 ## xmm3 = jxO 0 jxH1 jxH2 shufps $228,%xmm6,%xmm4 ## constant 11100100 ;# xmm4 = jyO 0 jyH1 jyH2 shufps $68,%xmm7,%xmm5 ## constant 01000100 ;# xmm5 = jzO 0 jzH1 jzH2 ## store all j coordinates in jO movaps %xmm3,nb302_jxO(%esp) movaps %xmm4,nb302_jyO(%esp) movaps %xmm5,nb302_jzO(%esp) subps %xmm3,%xmm0 subps %xmm4,%xmm1 subps %xmm5,%xmm2 movaps %xmm0,nb302_dxOO(%esp) movaps %xmm1,nb302_dyOO(%esp) movaps %xmm2,nb302_dzOO(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 ## have rsq in xmm0 ## do invsqrt rsqrtps %xmm0,%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb302_three(%esp),%xmm3 mulps %xmm0,%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb302_half(%esp),%xmm3 ## rinv iO - j water movaps %xmm3,%xmm1 mulps %xmm0,%xmm1 ## xmm1=r movaps %xmm3,%xmm0 ## xmm0=rinv 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,%ebx movd %mm7,%ecx psrlq $32,%mm7 movd %mm7,%edx ## table indices in ebx,ecx,edx movl nb302_VFtab(%ebp),%esi movlps (%esi,%ebx,4),%xmm5 movlps (%esi,%ecx,4),%xmm7 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,%ebx,4),%xmm7 movlps 8(%esi,%ecx,4),%xmm3 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 xorps %xmm3,%xmm3 ## fetch charges to xmm3 (temporary) movss nb302_qqOO(%esp),%xmm3 movhps nb302_qqOH(%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 xmm5 contains vcoul and xmm3 fijC addps nb302_vctot(%esp),%xmm5 movaps %xmm5,nb302_vctot(%esp) xorps %xmm2,%xmm2 mulps nb302_tsc(%esp),%xmm3 subps %xmm3,%xmm2 mulps %xmm2,%xmm0 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 mulps nb302_dxOO(%esp),%xmm0 mulps nb302_dyOO(%esp),%xmm1 mulps nb302_dzOO(%esp),%xmm2 ## initial update for j forces xorps %xmm3,%xmm3 xorps %xmm4,%xmm4 xorps %xmm5,%xmm5 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 movaps %xmm3,nb302_fjxO(%esp) movaps %xmm4,nb302_fjyO(%esp) movaps %xmm5,nb302_fjzO(%esp) addps nb302_fixO(%esp),%xmm0 addps nb302_fiyO(%esp),%xmm1 addps nb302_fizO(%esp),%xmm2 movaps %xmm0,nb302_fixO(%esp) movaps %xmm1,nb302_fiyO(%esp) movaps %xmm2,nb302_fizO(%esp) ## done with i O Now do i H1 & H2 simultaneously first get i particle coords: movaps nb302_ixH1(%esp),%xmm0 movaps nb302_iyH1(%esp),%xmm1 movaps nb302_izH1(%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_jxO(%esp),%xmm3 subps nb302_jyO(%esp),%xmm4 subps nb302_jzO(%esp),%xmm5 movaps %xmm0,nb302_dxH1O(%esp) movaps %xmm1,nb302_dyH1O(%esp) movaps %xmm2,nb302_dzH1O(%esp) movaps %xmm3,nb302_dxH2O(%esp) movaps %xmm4,nb302_dyH2O(%esp) movaps %xmm5,nb302_dzH2O(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm3,%xmm4 addps %xmm2,%xmm0 ## have rsqH1 in xmm0 addps %xmm5,%xmm4 ## have rsqH2 in xmm4 ## start with H1, save H2 data movaps %xmm4,nb302_rsqH2O(%esp) ## do invsqrt 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 ## rinv H1 - j water mulps nb302_half(%esp),%xmm7 ## rinv H2 - j water ## start with H1, save H2 data movaps %xmm7,nb302_rinvH2O(%esp) movaps %xmm3,%xmm1 mulps %xmm0,%xmm1 ## xmm1=r movaps %xmm3,%xmm0 ## xmm0=rinv 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,%ebx movd %mm7,%ecx psrlq $32,%mm7 movd %mm7,%edx ## table indices in ebx,ecx,edx movlps (%esi,%ebx,4),%xmm5 movlps (%esi,%ecx,4),%xmm7 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,%ebx,4),%xmm7 movlps 8(%esi,%ecx,4),%xmm3 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 xorps %xmm3,%xmm3 ## fetch charges to xmm3 (temporary) movss nb302_qqOH(%esp),%xmm3 movhps nb302_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 xmm5 contains vcoul and xmm3 fijC addps nb302_vctot(%esp),%xmm5 movaps %xmm5,nb302_vctot(%esp) xorps %xmm1,%xmm1 mulps nb302_tsc(%esp),%xmm3 mulps %xmm0,%xmm3 subps %xmm3,%xmm1 movaps %xmm1,%xmm0 movaps %xmm1,%xmm2 mulps nb302_dxH1O(%esp),%xmm0 mulps nb302_dyH1O(%esp),%xmm1 mulps nb302_dzH1O(%esp),%xmm2 ## update forces H1 - j water movaps nb302_fjxO(%esp),%xmm3 movaps nb302_fjyO(%esp),%xmm4 movaps nb302_fjzO(%esp),%xmm5 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 movaps %xmm3,nb302_fjxO(%esp) movaps %xmm4,nb302_fjyO(%esp) movaps %xmm5,nb302_fjzO(%esp) addps nb302_fixH1(%esp),%xmm0 addps nb302_fiyH1(%esp),%xmm1 addps nb302_fizH1(%esp),%xmm2 movaps %xmm0,nb302_fixH1(%esp) movaps %xmm1,nb302_fiyH1(%esp) movaps %xmm2,nb302_fizH1(%esp) ## do table for H2 - j water interaction movaps nb302_rinvH2O(%esp),%xmm0 movaps nb302_rsqH2O(%esp),%xmm1 mulps %xmm0,%xmm1 ## xmm0=rinv, 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,%ebx movd %mm7,%ecx psrlq $32,%mm7 movd %mm7,%edx ## table indices in ebx,ecx,edx movlps (%esi,%ebx,4),%xmm5 movlps (%esi,%ecx,4),%xmm7 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,%ebx,4),%xmm7 movlps 8(%esi,%ecx,4),%xmm3 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 xorps %xmm3,%xmm3 ## fetch charges to xmm3 (temporary) movss nb302_qqOH(%esp),%xmm3 movhps nb302_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 xmm5 contains vcoul and xmm3 fijC addps nb302_vctot(%esp),%xmm5 movaps %xmm5,nb302_vctot(%esp) xorps %xmm1,%xmm1 mulps nb302_tsc(%esp),%xmm3 mulps %xmm0,%xmm3 subps %xmm3,%xmm1 movaps %xmm1,%xmm0 movaps %xmm1,%xmm2 mulps nb302_dxH2O(%esp),%xmm0 mulps nb302_dyH2O(%esp),%xmm1 mulps nb302_dzH2O(%esp),%xmm2 movaps nb302_fjxO(%esp),%xmm3 movaps nb302_fjyO(%esp),%xmm4 movaps nb302_fjzO(%esp),%xmm5 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 movl nb302_faction(%ebp),%esi movaps %xmm3,nb302_fjxO(%esp) movaps %xmm4,nb302_fjyO(%esp) movaps %xmm5,nb302_fjzO(%esp) addps nb302_fixH2(%esp),%xmm0 addps nb302_fiyH2(%esp),%xmm1 addps nb302_fizH2(%esp),%xmm2 movaps %xmm0,nb302_fixH2(%esp) movaps %xmm1,nb302_fiyH2(%esp) movaps %xmm2,nb302_fizH2(%esp) ## update j water forces from local variables movlps (%esi,%eax,4),%xmm0 movlps 12(%esi,%eax,4),%xmm1 movhps 24(%esi,%eax,4),%xmm1 movaps nb302_fjxO(%esp),%xmm3 movaps nb302_fjyO(%esp),%xmm4 movaps nb302_fjzO(%esp),%xmm5 movaps %xmm5,%xmm6 movaps %xmm5,%xmm7 shufps $2,%xmm6,%xmm6 ## constant 00000010 shufps $3,%xmm7,%xmm7 ## constant 00000011 addss 8(%esi,%eax,4),%xmm5 addss 20(%esi,%eax,4),%xmm6 addss 32(%esi,%eax,4),%xmm7 movss %xmm5,8(%esi,%eax,4) movss %xmm6,20(%esi,%eax,4) movss %xmm7,32(%esi,%eax,4) movaps %xmm3,%xmm5 unpcklps %xmm4,%xmm3 unpckhps %xmm4,%xmm5 addps %xmm3,%xmm0 addps %xmm5,%xmm1 movlps %xmm0,(%esi,%eax,4) movlps %xmm1,12(%esi,%eax,4) movhps %xmm1,24(%esi,%eax,4) decl nb302_innerk(%esp) jz _nb_kernel302_ia32_sse.nb302_updateouterdata jmp _nb_kernel302_ia32_sse.nb302_single_loop_nb_kernel302_ia32_sse.nb302_updateouterdata: movl nb302_ii3(%esp),%ecx⌨️ 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?