nb_kernel112_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,177 行 · 第 1/5 页
S
2,177 行
;# xmm6= jxH1b jyH1b jxH1d jyH1d ;# xmm7= jxH2b jyH2b jxH2d jyH2d movaps xmm0, xmm2 movaps xmm1, xmm3 unpcklps xmm0, xmm5 ;# xmm0= jxOa jxOb jyOa jyOb unpcklps xmm1, xmm6 ;# xmm1= jxH1a jxH1b jyH1a jyH1b unpckhps xmm2, xmm5 ;# xmm2= jxOc jxOd jyOc jyOd unpckhps xmm3, xmm6 ;# xmm3= jxH1c jxH1d jyH1c jyH1d movaps xmm5, xmm4 movaps xmm6, xmm0 unpcklps xmm4, xmm7 ;# xmm4= jxH2a jxH2b jyH2a jyH2b unpckhps xmm5, xmm7 ;# xmm5= jxH2c jxH2d jyH2c jyH2d movaps xmm7, xmm1 movlhps xmm0, xmm2 ;# xmm0= jxOa jxOb jxOc jxOd movaps [esp + nb112_jxO], xmm0 movhlps xmm2, xmm6 ;# xmm2= jyOa jyOb jyOc jyOd movaps [esp + nb112_jyO], xmm2 movlhps xmm1, xmm3 movaps [esp + nb112_jxH1], xmm1 movhlps xmm3, xmm7 movaps xmm6, xmm4 movaps [esp + nb112_jyH1], xmm3 movlhps xmm4, xmm5 movaps [esp + nb112_jxH2], xmm4 movhlps xmm5, xmm6 movaps [esp + nb112_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 + nb112_jzO], xmm0 movaps [esp + nb112_jzH1], xmm1 movaps [esp + nb112_jzH2], xmm2 movaps xmm0, [esp + nb112_ixO] movaps xmm1, [esp + nb112_iyO] movaps xmm2, [esp + nb112_izO] movaps xmm3, [esp + nb112_ixO] movaps xmm4, [esp + nb112_iyO] movaps xmm5, [esp + nb112_izO] subps xmm0, [esp + nb112_jxO] subps xmm1, [esp + nb112_jyO] subps xmm2, [esp + nb112_jzO] subps xmm3, [esp + nb112_jxH1] subps xmm4, [esp + nb112_jyH1] subps xmm5, [esp + nb112_jzH1] movaps [esp + nb112_dxOO], xmm0 movaps [esp + nb112_dyOO], xmm1 movaps [esp + nb112_dzOO], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb112_dxOH1], xmm3 movaps [esp + nb112_dyOH1], xmm4 movaps [esp + nb112_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 + nb112_rsqOO], xmm0 movaps [esp + nb112_rsqOH1], xmm3 movaps xmm0, [esp + nb112_ixO] movaps xmm1, [esp + nb112_iyO] movaps xmm2, [esp + nb112_izO] movaps xmm3, [esp + nb112_ixH1] movaps xmm4, [esp + nb112_iyH1] movaps xmm5, [esp + nb112_izH1] subps xmm0, [esp + nb112_jxH2] subps xmm1, [esp + nb112_jyH2] subps xmm2, [esp + nb112_jzH2] subps xmm3, [esp + nb112_jxO] subps xmm4, [esp + nb112_jyO] subps xmm5, [esp + nb112_jzO] movaps [esp + nb112_dxOH2], xmm0 movaps [esp + nb112_dyOH2], xmm1 movaps [esp + nb112_dzOH2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb112_dxH1O], xmm3 movaps [esp + nb112_dyH1O], xmm4 movaps [esp + nb112_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 + nb112_rsqOH2], xmm0 movaps [esp + nb112_rsqH1O], xmm3 movaps xmm0, [esp + nb112_ixH1] movaps xmm1, [esp + nb112_iyH1] movaps xmm2, [esp + nb112_izH1] movaps xmm3, [esp + nb112_ixH1] movaps xmm4, [esp + nb112_iyH1] movaps xmm5, [esp + nb112_izH1] subps xmm0, [esp + nb112_jxH1] subps xmm1, [esp + nb112_jyH1] subps xmm2, [esp + nb112_jzH1] subps xmm3, [esp + nb112_jxH2] subps xmm4, [esp + nb112_jyH2] subps xmm5, [esp + nb112_jzH2] movaps [esp + nb112_dxH1H1], xmm0 movaps [esp + nb112_dyH1H1], xmm1 movaps [esp + nb112_dzH1H1], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb112_dxH1H2], xmm3 movaps [esp + nb112_dyH1H2], xmm4 movaps [esp + nb112_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 + nb112_rsqH1H1], xmm0 movaps [esp + nb112_rsqH1H2], xmm3 movaps xmm0, [esp + nb112_ixH2] movaps xmm1, [esp + nb112_iyH2] movaps xmm2, [esp + nb112_izH2] movaps xmm3, [esp + nb112_ixH2] movaps xmm4, [esp + nb112_iyH2] movaps xmm5, [esp + nb112_izH2] subps xmm0, [esp + nb112_jxO] subps xmm1, [esp + nb112_jyO] subps xmm2, [esp + nb112_jzO] subps xmm3, [esp + nb112_jxH1] subps xmm4, [esp + nb112_jyH1] subps xmm5, [esp + nb112_jzH1] movaps [esp + nb112_dxH2O], xmm0 movaps [esp + nb112_dyH2O], xmm1 movaps [esp + nb112_dzH2O], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb112_dxH2H1], xmm3 movaps [esp + nb112_dyH2H1], xmm4 movaps [esp + nb112_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 + nb112_rsqH2O], xmm0 movaps [esp + nb112_rsqH2H1], xmm4 movaps xmm0, [esp + nb112_ixH2] movaps xmm1, [esp + nb112_iyH2] movaps xmm2, [esp + nb112_izH2] subps xmm0, [esp + nb112_jxH2] subps xmm1, [esp + nb112_jyH2] subps xmm2, [esp + nb112_jzH2] movaps [esp + nb112_dxH2H2], xmm0 movaps [esp + nb112_dyH2H2], xmm1 movaps [esp + nb112_dzH2H2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 movaps [esp + nb112_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 + nb112_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 + nb112_half] ;# rinvH2H2 mulps xmm7, [esp + nb112_half] ;# rinvH2H1 movaps [esp + nb112_rinvH2H2], xmm3 movaps [esp + nb112_rinvH2H1], xmm7 rsqrtps xmm1, [esp + nb112_rsqOO] rsqrtps xmm5, [esp + nb112_rsqOH1] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb112_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb112_rsqOO] mulps xmm5, [esp + nb112_rsqOH1] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb112_half] mulps xmm7, [esp + nb112_half] movaps [esp + nb112_rinvOO], xmm3 movaps [esp + nb112_rinvOH1], xmm7 rsqrtps xmm1, [esp + nb112_rsqOH2] rsqrtps xmm5, [esp + nb112_rsqH1O] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb112_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb112_rsqOH2] mulps xmm5, [esp + nb112_rsqH1O] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb112_half] mulps xmm7, [esp + nb112_half] movaps [esp + nb112_rinvOH2], xmm3 movaps [esp + nb112_rinvH1O], xmm7 rsqrtps xmm1, [esp + nb112_rsqH1H1] rsqrtps xmm5, [esp + nb112_rsqH1H2] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb112_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb112_rsqH1H1] mulps xmm5, [esp + nb112_rsqH1H2] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb112_half] mulps xmm7, [esp + nb112_half] movaps [esp + nb112_rinvH1H1], xmm3 movaps [esp + nb112_rinvH1H2], xmm7 rsqrtps xmm1, [esp + nb112_rsqH2O] movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [esp + nb112_three] mulps xmm1, [esp + nb112_rsqH2O] subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [esp + nb112_half] movaps [esp + nb112_rinvH2O], xmm3 ;# start with OO interaction movaps xmm0, [esp + nb112_rinvOO] movaps xmm7, xmm0 mulps xmm0, xmm0 movaps xmm1, xmm0 mulps xmm1, xmm0 mulps xmm1, xmm0 ;# xmm1=rinvsix mulps xmm7, [esp + nb112_qqOO] movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=rinvtwelve mulps xmm1, [esp + nb112_c6] mulps xmm2, [esp + nb112_c12] movaps xmm3, xmm2 subps xmm3, xmm1 ;# xmm3=Vvdw12-Vvdw6 addps xmm3, [esp + nb112_Vvdwtot] mulps xmm1, [esp + nb112_six] mulps xmm2, [esp + nb112_twelve] movaps [esp + nb112_Vvdwtot], xmm3 subps xmm2, xmm1 addps xmm2, xmm7 addps xmm7, [esp + nb112_vctot] mulps xmm0, xmm2 movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb112_dxOO] mulps xmm1, [esp + nb112_dyOO] mulps xmm2, [esp + nb112_dzOO] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb112_fixO] addps xmm1, [esp + nb112_fiyO] addps xmm2, [esp + nb112_fizO] movaps [esp + nb112_fjxO], xmm3 movaps [esp + nb112_fjyO], xmm4 movaps [esp + nb112_fjzO], xmm5 movaps [esp + nb112_fixO], xmm0 movaps [esp + nb112_fiyO], xmm1 movaps [esp + nb112_fizO], xmm2 ;# O-H1 interaction movaps xmm0, [esp + nb112_rinvOH1] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb112_qqOH] mulps xmm0, xmm1 ;# fsOH1 addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb112_dxOH1] mulps xmm1, [esp + nb112_dyOH1] mulps xmm2, [esp + nb112_dzOH1] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb112_fixO] addps xmm1, [esp + nb112_fiyO] addps xmm2, [esp + nb112_fizO] movaps [esp + nb112_fjxH1], xmm3 movaps [esp + nb112_fjyH1], xmm4 movaps [esp + nb112_fjzH1], xmm5 movaps [esp + nb112_fixO], xmm0 movaps [esp + nb112_fiyO], xmm1 movaps [esp + nb112_fizO], xmm2 ;# O-H2 interaction movaps xmm0, [esp + nb112_rinvOH2] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb112_qqOH] mulps xmm0, xmm1 ;# fsOH2 addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb112_dxOH2] mulps xmm1, [esp + nb112_dyOH2] mulps xmm2, [esp + nb112_dzOH2] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb112_fixO] addps xmm1, [esp + nb112_fiyO] addps xmm2, [esp + nb112_fizO] movaps [esp + nb112_fjxH2], xmm3 movaps [esp + nb112_fjyH2], xmm4 movaps [esp + nb112_fjzH2], xmm5 movaps [esp + nb112_fixO], xmm0 movaps [esp + nb112_fiyO], xmm1 movaps [esp + nb112_fizO], xmm2 ;# H1-O interaction movaps xmm0, [esp + nb112_rinvH1O] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb112_qqOH] mulps xmm0, xmm1 ;# fsH1O addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps xmm2, xmm0 movaps xmm3, [esp + nb112_fjxO] movaps xmm4, [esp + nb112_fjyO] movaps xmm5, [esp + nb112_fjzO] mulps xmm0, [esp + nb112_dxH1O] mulps xmm1, [esp + nb112_dyH1O] mulps xmm2, [esp + nb112_dzH1O] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb112_fixH1] addps xmm1, [esp + nb112_fiyH1] addps xmm2, [esp + nb112_fizH1] movaps [esp + nb112_fjxO], xmm3 movaps [esp + nb112_fjyO], xmm4 movaps [esp + nb112_fjzO], xmm5 movaps [esp + nb112_fixH1], xmm0 movaps [esp + nb112_fiyH1], xmm1 movaps [esp + nb112_fizH1], xmm2 ;# H1-H1 interaction movaps xmm0, [esp + nb112_rinvH1H1] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb112_qqHH] mulps xmm0, xmm1 ;# fsH1H1 addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps xmm2, xmm0 movaps xmm3, [esp + nb112_fjxH1] movaps xmm4, [esp + nb112_fjyH1] movaps xmm5, [esp + nb112_fjzH1] mulps xmm0, [esp + nb112_dxH1H1] mulps xmm1, [esp + nb112_dyH1H1] mulps xmm2, [esp + nb112_dzH1H1] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb112_fixH1] addps xmm1, [esp + nb112_fiyH1] addps xmm2, [esp + nb112_fizH1] movaps [esp + nb112_fjxH1], xmm3 movaps [esp + nb112_fjyH1], xmm4 movaps [esp + nb112_fjzH1], xmm5 movaps [esp + nb112_fixH1], xmm0 movaps [esp + nb112_fiyH1], xmm1 movaps [esp + nb112_fizH1], xmm2
⌨️ 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?