nb_kernel202_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,166 行 · 第 1/5 页
S
2,166 行
movaps xmm7, xmm1 movlhps xmm0, xmm2 ;# xmm0= jxOa jxOb jxOc jxOd movaps [esp + nb202_jxO], xmm0 movhlps xmm2, xmm6 ;# xmm2= jyOa jyOb jyOc jyOd movaps [esp + nb202_jyO], xmm2 movlhps xmm1, xmm3 movaps [esp + nb202_jxH1], xmm1 movhlps xmm3, xmm7 movaps xmm6, xmm4 movaps [esp + nb202_jyH1], xmm3 movlhps xmm4, xmm5 movaps [esp + nb202_jxH2], xmm4 movhlps xmm5, xmm6 movaps [esp + nb202_jyH2], xmm5 movss xmm0, [esi + eax*4 + 8] movss xmm1, [esi + eax*4 + 20] movss xmm2, [esi + eax*4 + 32] movss xmm3, [esi + ecx*4 + 8] movss xmm4, [esi + ecx*4 + 20] movss xmm5, [esi + ecx*4 + 32] movhps xmm0, [esi + ebx*4 + 4] movhps xmm1, [esi + ebx*4 + 16] movhps xmm2, [esi + ebx*4 + 28] movhps xmm3, [esi + edx*4 + 4] movhps xmm4, [esi + edx*4 + 16] movhps xmm5, [esi + edx*4 + 28] shufps xmm0, xmm3, 204 ;# constant 11001100 shufps xmm1, xmm4, 204 ;# constant 11001100 shufps xmm2, xmm5, 204 ;# constant 11001100 movaps [esp + nb202_jzO], xmm0 movaps [esp + nb202_jzH1], xmm1 movaps [esp + nb202_jzH2], xmm2 movaps xmm0, [esp + nb202_ixO] movaps xmm1, [esp + nb202_iyO] movaps xmm2, [esp + nb202_izO] movaps xmm3, [esp + nb202_ixO] movaps xmm4, [esp + nb202_iyO] movaps xmm5, [esp + nb202_izO] subps xmm0, [esp + nb202_jxO] subps xmm1, [esp + nb202_jyO] subps xmm2, [esp + nb202_jzO] subps xmm3, [esp + nb202_jxH1] subps xmm4, [esp + nb202_jyH1] subps xmm5, [esp + nb202_jzH1] movaps [esp + nb202_dxOO], xmm0 movaps [esp + nb202_dyOO], xmm1 movaps [esp + nb202_dzOO], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb202_dxOH1], xmm3 movaps [esp + nb202_dyOH1], xmm4 movaps [esp + nb202_dzOH1], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm3, xmm4 addps xmm3, xmm5 movaps [esp + nb202_rsqOO], xmm0 movaps [esp + nb202_rsqOH1], xmm3 movaps xmm0, [esp + nb202_ixO] movaps xmm1, [esp + nb202_iyO] movaps xmm2, [esp + nb202_izO] movaps xmm3, [esp + nb202_ixH1] movaps xmm4, [esp + nb202_iyH1] movaps xmm5, [esp + nb202_izH1] subps xmm0, [esp + nb202_jxH2] subps xmm1, [esp + nb202_jyH2] subps xmm2, [esp + nb202_jzH2] subps xmm3, [esp + nb202_jxO] subps xmm4, [esp + nb202_jyO] subps xmm5, [esp + nb202_jzO] movaps [esp + nb202_dxOH2], xmm0 movaps [esp + nb202_dyOH2], xmm1 movaps [esp + nb202_dzOH2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb202_dxH1O], xmm3 movaps [esp + nb202_dyH1O], xmm4 movaps [esp + nb202_dzH1O], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm3, xmm4 addps xmm3, xmm5 movaps [esp + nb202_rsqOH2], xmm0 movaps [esp + nb202_rsqH1O], xmm3 movaps xmm0, [esp + nb202_ixH1] movaps xmm1, [esp + nb202_iyH1] movaps xmm2, [esp + nb202_izH1] movaps xmm3, [esp + nb202_ixH1] movaps xmm4, [esp + nb202_iyH1] movaps xmm5, [esp + nb202_izH1] subps xmm0, [esp + nb202_jxH1] subps xmm1, [esp + nb202_jyH1] subps xmm2, [esp + nb202_jzH1] subps xmm3, [esp + nb202_jxH2] subps xmm4, [esp + nb202_jyH2] subps xmm5, [esp + nb202_jzH2] movaps [esp + nb202_dxH1H1], xmm0 movaps [esp + nb202_dyH1H1], xmm1 movaps [esp + nb202_dzH1H1], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb202_dxH1H2], xmm3 movaps [esp + nb202_dyH1H2], xmm4 movaps [esp + nb202_dzH1H2], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm3, xmm4 addps xmm3, xmm5 movaps [esp + nb202_rsqH1H1], xmm0 movaps [esp + nb202_rsqH1H2], xmm3 movaps xmm0, [esp + nb202_ixH2] movaps xmm1, [esp + nb202_iyH2] movaps xmm2, [esp + nb202_izH2] movaps xmm3, [esp + nb202_ixH2] movaps xmm4, [esp + nb202_iyH2] movaps xmm5, [esp + nb202_izH2] subps xmm0, [esp + nb202_jxO] subps xmm1, [esp + nb202_jyO] subps xmm2, [esp + nb202_jzO] subps xmm3, [esp + nb202_jxH1] subps xmm4, [esp + nb202_jyH1] subps xmm5, [esp + nb202_jzH1] movaps [esp + nb202_dxH2O], xmm0 movaps [esp + nb202_dyH2O], xmm1 movaps [esp + nb202_dzH2O], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb202_dxH2H1], xmm3 movaps [esp + nb202_dyH2H1], xmm4 movaps [esp + nb202_dzH2H1], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm4, xmm3 addps xmm4, xmm5 movaps [esp + nb202_rsqH2O], xmm0 movaps [esp + nb202_rsqH2H1], xmm4 movaps xmm0, [esp + nb202_ixH2] movaps xmm1, [esp + nb202_iyH2] movaps xmm2, [esp + nb202_izH2] subps xmm0, [esp + nb202_jxH2] subps xmm1, [esp + nb202_jyH2] subps xmm2, [esp + nb202_jzH2] movaps [esp + nb202_dxH2H2], xmm0 movaps [esp + nb202_dyH2H2], xmm1 movaps [esp + nb202_dzH2H2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 movaps [esp + nb202_rsqH2H2], xmm0 ;# start doing invsqrt use rsq values in xmm0, xmm4 rsqrtps xmm1, xmm0 rsqrtps xmm5, xmm4 movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb202_three] movaps xmm7, xmm3 mulps xmm1, xmm0 mulps xmm5, xmm4 subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb202_half] ;# rinvH2H2 mulps xmm7, [esp + nb202_half] ;# rinvH2H1 movaps [esp + nb202_rinvH2H2], xmm3 movaps [esp + nb202_rinvH2H1], xmm7 rsqrtps xmm1, [esp + nb202_rsqOO] rsqrtps xmm5, [esp + nb202_rsqOH1] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb202_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb202_rsqOO] mulps xmm5, [esp + nb202_rsqOH1] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb202_half] mulps xmm7, [esp + nb202_half] movaps [esp + nb202_rinvOO], xmm3 movaps [esp + nb202_rinvOH1], xmm7 rsqrtps xmm1, [esp + nb202_rsqOH2] rsqrtps xmm5, [esp + nb202_rsqH1O] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb202_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb202_rsqOH2] mulps xmm5, [esp + nb202_rsqH1O] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb202_half] mulps xmm7, [esp + nb202_half] movaps [esp + nb202_rinvOH2], xmm3 movaps [esp + nb202_rinvH1O], xmm7 rsqrtps xmm1, [esp + nb202_rsqH1H1] rsqrtps xmm5, [esp + nb202_rsqH1H2] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb202_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb202_rsqH1H1] mulps xmm5, [esp + nb202_rsqH1H2] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb202_half] mulps xmm7, [esp + nb202_half] movaps [esp + nb202_rinvH1H1], xmm3 movaps [esp + nb202_rinvH1H2], xmm7 rsqrtps xmm1, [esp + nb202_rsqH2O] movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [esp + nb202_three] mulps xmm1, [esp + nb202_rsqH2O] subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [esp + nb202_half] movaps [esp + nb202_rinvH2O], xmm3 ;# start with OO interaction movaps xmm0, [esp + nb202_rinvOO] movaps xmm7, xmm0 ;# xmm7=rinv movaps xmm5, [esp + nb202_krf] mulps xmm0, xmm0 ;# xmm0=rinvsq mulps xmm5, [esp + nb202_rsqOO] ;# xmm5=krsq movaps xmm6, xmm5 addps xmm6, xmm7 ;# xmm6=rinv+ krsq subps xmm6, [esp + nb202_crf] mulps xmm6, [esp + nb202_qqOO] ;# xmm6=voul=qq*(rinv+ krsq-crf) mulps xmm5, [esp + nb202_two] subps xmm7, xmm5 ;# xmm7=rinv-2*krsq mulps xmm7, [esp + nb202_qqOO] ;# xmm7 = coul part of fscal addps xmm6, [esp + nb202_vctot] ;# local vctot summation variable mulps xmm0, xmm7 movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb202_dxOO] mulps xmm1, [esp + nb202_dyOO] mulps xmm2, [esp + nb202_dzOO] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb202_fixO] addps xmm1, [esp + nb202_fiyO] addps xmm2, [esp + nb202_fizO] movaps [esp + nb202_fjxO], xmm3 movaps [esp + nb202_fjyO], xmm4 movaps [esp + nb202_fjzO], xmm5 movaps [esp + nb202_fixO], xmm0 movaps [esp + nb202_fiyO], xmm1 movaps [esp + nb202_fizO], xmm2 ;# O-H1 interaction movaps xmm0, [esp + nb202_rinvOH1] movaps xmm7, xmm0 ;# xmm7=rinv movaps xmm5, [esp + nb202_krf] movaps xmm1, xmm0 mulps xmm5, [esp + nb202_rsqOH1] ;# xmm5=krsq movaps xmm4, xmm5 addps xmm4, xmm7 ;# xmm4=r inv+ krsq subps xmm4, [esp + nb202_crf] mulps xmm0, xmm0 mulps xmm4, [esp + nb202_qqOH] ;# xmm4=voul=qq*(rinv+ krsq-crf) mulps xmm5, [esp + nb202_two] subps xmm7, xmm5 ;# xmm7=rinv-2*krsq mulps xmm7, [esp + nb202_qqOH] ;# xmm7 = coul part of fscal addps xmm6, xmm4 ;# add to local vctot mulps xmm0, xmm7 ;# fsOH1 movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb202_dxOH1] mulps xmm1, [esp + nb202_dyOH1] mulps xmm2, [esp + nb202_dzOH1] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb202_fixO] addps xmm1, [esp + nb202_fiyO] addps xmm2, [esp + nb202_fizO] movaps [esp + nb202_fjxH1], xmm3 movaps [esp + nb202_fjyH1], xmm4 movaps [esp + nb202_fjzH1], xmm5 movaps [esp + nb202_fixO], xmm0 movaps [esp + nb202_fiyO], xmm1 movaps [esp + nb202_fizO], xmm2 ;# O-H2 interaction movaps xmm0, [esp + nb202_rinvOH2] movaps xmm7, xmm0 ;# xmm7=rinv movaps xmm5, [esp + nb202_krf] movaps xmm1, xmm0 mulps xmm5, [esp + nb202_rsqOH2] ;# xmm5=krsq movaps xmm4, xmm5 addps xmm4, xmm7 ;# xmm4=r inv+ krsq subps xmm4, [esp + nb202_crf] mulps xmm0, xmm0 mulps xmm4, [esp + nb202_qqOH] ;# xmm4=voul=qq*(rinv+ krsq-crf) mulps xmm5, [esp + nb202_two] subps xmm7, xmm5 ;# xmm7=rinv-2*krsq mulps xmm7, [esp + nb202_qqOH] ;# xmm7 = coul part of fscal addps xmm6, xmm4 ;# add to local vctot mulps xmm0, xmm7 ;# fsOH2 movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb202_dxOH2] mulps xmm1, [esp + nb202_dyOH2] mulps xmm2, [esp + nb202_dzOH2] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb202_fixO] addps xmm1, [esp + nb202_fiyO] addps xmm2, [esp + nb202_fizO] movaps [esp + nb202_fjxH2], xmm3 movaps [esp + nb202_fjyH2], xmm4 movaps [esp + nb202_fjzH2], xmm5 movaps [esp + nb202_fixO], xmm0 movaps [esp + nb202_fiyO], xmm1 movaps [esp + nb202_fizO], xmm2 ;# H1-O interaction movaps xmm0, [esp + nb202_rinvH1O] movaps xmm7, xmm0 ;# xmm7=rinv movaps xmm5, [esp + nb202_krf] movaps xmm1, xmm0 mulps xmm5, [esp + nb202_rsqH1O] ;# xmm5=krsq movaps xmm4, xmm5 addps xmm4, xmm7 ;# xmm4=r inv+ krsq subps xmm4, [esp + nb202_crf] mulps xmm0, xmm0 mulps xmm4, [esp + nb202_qqOH] ;# xmm4=voul=qq*(rinv+ krsq-crf) mulps xmm5, [esp + nb202_two] subps xmm7, xmm5 ;# xmm7=rinv-2*krsq mulps xmm7, [esp + nb202_qqOH] ;# xmm7 = coul part of fscal addps xmm6, xmm4 ;# add to local vctot mulps xmm0, xmm7 ;# fsOH2 movaps xmm1, xmm0 movaps xmm2, xmm0 movaps xmm3, [esp + nb202_fjxO] movaps xmm4, [esp + nb202_fjyO] movaps xmm5, [esp + nb202_fjzO] mulps xmm0, [esp + nb202_dxH1O] mulps xmm1, [esp + nb202_dyH1O] mulps xmm2, [esp + nb202_dzH1O] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb202_fixH1] addps xmm1, [esp + nb202_fiyH1] addps xmm2, [esp + nb202_fizH1] movaps [esp + nb202_fjxO], xmm3 movaps [esp + nb202_fjyO], xmm4 movaps [esp + nb202_fjzO], xmm5 movaps [esp + nb202_fixH1], xmm0 movaps [esp + nb202_fiyH1], xmm1 movaps [esp + nb202_fizH1], xmm2 ;# H1-H1 interaction movaps xmm0, [esp + nb202_rinvH1H1] movaps xmm7, xmm0 ;# xmm7=rinv movaps xmm5, [esp + nb202_krf] movaps xmm1, xmm0 mulps xmm5, [esp + nb202_rsqH1H1] ;# xmm5=krsq movaps xmm4, xmm5 addps xmm4, xmm7 ;# xmm4=r inv+ krsq subps xmm4, [esp + nb202_crf] mulps xmm0, xmm0 mulps xmm4, [esp + nb202_qqHH] ;# xmm4=voul=qq*(rinv+ krsq-crf) mulps xmm5, [esp + nb202_two] subps xmm7, xmm5 ;# xmm7=rinv-2*krsq mulps xmm7, [esp + nb202_qqHH] ;# xmm7 = coul part of fscal addps xmm6, xmm4 ;# add to local vctot
⌨️ 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?