nb_kernel212_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,151 行 · 第 1/5 页
S
2,151 行
movlps xmm5, [esi + ebx*4] movlps xmm6, [esi + ebx*4 + 12] movlps xmm7, [esi + ebx*4 + 24] movhps xmm2, [esi + ecx*4] movhps xmm3, [esi + ecx*4 + 12] movhps xmm4, [esi + ecx*4 + 24] movhps xmm5, [esi + edx*4] movhps xmm6, [esi + edx*4 + 12] movhps xmm7, [esi + edx*4 + 24] ;# current state: ;# xmm2= jxOa jyOa jxOc jyOc ;# xmm3= jxH1a jyH1a jxH1c jyH1c ;# xmm4= jxH2a jyH2a jxH2c jyH2c ;# xmm5= jxOb jyOb jxOd jyOd ;# 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 + nb212_jxO], xmm0 movhlps xmm2, xmm6 ;# xmm2= jyOa jyOb jyOc jyOd movaps [esp + nb212_jyO], xmm2 movlhps xmm1, xmm3 movaps [esp + nb212_jxH1], xmm1 movhlps xmm3, xmm7 movaps xmm6, xmm4 movaps [esp + nb212_jyH1], xmm3 movlhps xmm4, xmm5 movaps [esp + nb212_jxH2], xmm4 movhlps xmm5, xmm6 movaps [esp + nb212_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 + nb212_jzO], xmm0 movaps [esp + nb212_jzH1], xmm1 movaps [esp + nb212_jzH2], xmm2 movaps xmm0, [esp + nb212_ixO] movaps xmm1, [esp + nb212_iyO] movaps xmm2, [esp + nb212_izO] movaps xmm3, [esp + nb212_ixO] movaps xmm4, [esp + nb212_iyO] movaps xmm5, [esp + nb212_izO] subps xmm0, [esp + nb212_jxO] subps xmm1, [esp + nb212_jyO] subps xmm2, [esp + nb212_jzO] subps xmm3, [esp + nb212_jxH1] subps xmm4, [esp + nb212_jyH1] subps xmm5, [esp + nb212_jzH1] movaps [esp + nb212_dxOO], xmm0 movaps [esp + nb212_dyOO], xmm1 movaps [esp + nb212_dzOO], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb212_dxOH1], xmm3 movaps [esp + nb212_dyOH1], xmm4 movaps [esp + nb212_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 + nb212_rsqOO], xmm0 movaps [esp + nb212_rsqOH1], xmm3 movaps xmm0, [esp + nb212_ixO] movaps xmm1, [esp + nb212_iyO] movaps xmm2, [esp + nb212_izO] movaps xmm3, [esp + nb212_ixH1] movaps xmm4, [esp + nb212_iyH1] movaps xmm5, [esp + nb212_izH1] subps xmm0, [esp + nb212_jxH2] subps xmm1, [esp + nb212_jyH2] subps xmm2, [esp + nb212_jzH2] subps xmm3, [esp + nb212_jxO] subps xmm4, [esp + nb212_jyO] subps xmm5, [esp + nb212_jzO] movaps [esp + nb212_dxOH2], xmm0 movaps [esp + nb212_dyOH2], xmm1 movaps [esp + nb212_dzOH2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb212_dxH1O], xmm3 movaps [esp + nb212_dyH1O], xmm4 movaps [esp + nb212_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 + nb212_rsqOH2], xmm0 movaps [esp + nb212_rsqH1O], xmm3 movaps xmm0, [esp + nb212_ixH1] movaps xmm1, [esp + nb212_iyH1] movaps xmm2, [esp + nb212_izH1] movaps xmm3, [esp + nb212_ixH1] movaps xmm4, [esp + nb212_iyH1] movaps xmm5, [esp + nb212_izH1] subps xmm0, [esp + nb212_jxH1] subps xmm1, [esp + nb212_jyH1] subps xmm2, [esp + nb212_jzH1] subps xmm3, [esp + nb212_jxH2] subps xmm4, [esp + nb212_jyH2] subps xmm5, [esp + nb212_jzH2] movaps [esp + nb212_dxH1H1], xmm0 movaps [esp + nb212_dyH1H1], xmm1 movaps [esp + nb212_dzH1H1], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb212_dxH1H2], xmm3 movaps [esp + nb212_dyH1H2], xmm4 movaps [esp + nb212_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 + nb212_rsqH1H1], xmm0 movaps [esp + nb212_rsqH1H2], xmm3 movaps xmm0, [esp + nb212_ixH2] movaps xmm1, [esp + nb212_iyH2] movaps xmm2, [esp + nb212_izH2] movaps xmm3, [esp + nb212_ixH2] movaps xmm4, [esp + nb212_iyH2] movaps xmm5, [esp + nb212_izH2] subps xmm0, [esp + nb212_jxO] subps xmm1, [esp + nb212_jyO] subps xmm2, [esp + nb212_jzO] subps xmm3, [esp + nb212_jxH1] subps xmm4, [esp + nb212_jyH1] subps xmm5, [esp + nb212_jzH1] movaps [esp + nb212_dxH2O], xmm0 movaps [esp + nb212_dyH2O], xmm1 movaps [esp + nb212_dzH2O], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb212_dxH2H1], xmm3 movaps [esp + nb212_dyH2H1], xmm4 movaps [esp + nb212_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 + nb212_rsqH2O], xmm0 movaps [esp + nb212_rsqH2H1], xmm4 movaps xmm0, [esp + nb212_ixH2] movaps xmm1, [esp + nb212_iyH2] movaps xmm2, [esp + nb212_izH2] subps xmm0, [esp + nb212_jxH2] subps xmm1, [esp + nb212_jyH2] subps xmm2, [esp + nb212_jzH2] movaps [esp + nb212_dxH2H2], xmm0 movaps [esp + nb212_dyH2H2], xmm1 movaps [esp + nb212_dzH2H2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 movaps [esp + nb212_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 + nb212_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 + nb212_half] ;# rinvH2H2 mulps xmm7, [esp + nb212_half] ;# rinvH2H1 movaps [esp + nb212_rinvH2H2], xmm3 movaps [esp + nb212_rinvH2H1], xmm7 rsqrtps xmm1, [esp + nb212_rsqOO] rsqrtps xmm5, [esp + nb212_rsqOH1] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb212_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb212_rsqOO] mulps xmm5, [esp + nb212_rsqOH1] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb212_half] mulps xmm7, [esp + nb212_half] movaps [esp + nb212_rinvOO], xmm3 movaps [esp + nb212_rinvOH1], xmm7 rsqrtps xmm1, [esp + nb212_rsqOH2] rsqrtps xmm5, [esp + nb212_rsqH1O] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb212_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb212_rsqOH2] mulps xmm5, [esp + nb212_rsqH1O] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb212_half] mulps xmm7, [esp + nb212_half] movaps [esp + nb212_rinvOH2], xmm3 movaps [esp + nb212_rinvH1O], xmm7 rsqrtps xmm1, [esp + nb212_rsqH1H1] rsqrtps xmm5, [esp + nb212_rsqH1H2] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb212_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb212_rsqH1H1] mulps xmm5, [esp + nb212_rsqH1H2] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb212_half] mulps xmm7, [esp + nb212_half] movaps [esp + nb212_rinvH1H1], xmm3 movaps [esp + nb212_rinvH1H2], xmm7 rsqrtps xmm1, [esp + nb212_rsqH2O] movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [esp + nb212_three] mulps xmm1, [esp + nb212_rsqH2O] subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [esp + nb212_half] movaps [esp + nb212_rinvH2O], xmm3 ;# start with OO interaction movaps xmm0, [esp + nb212_rinvOO] movaps xmm7, xmm0 ;# xmm7=rinv movaps xmm5, [esp + nb212_krf] mulps xmm0, xmm0 movaps xmm1, xmm0 mulps xmm1, xmm0 mulps xmm1, xmm0 ;# xmm1=rinvsix mulps xmm5, [esp + nb212_rsqOO] ;# xmm5=krsq movaps xmm6, xmm5 addps xmm6, xmm7 ;# xmm6=rinv+ krsq subps xmm6, [esp + nb212_crf] mulps xmm6, [esp + nb212_qqOO] ;# xmm6=voul=qq*(rinv+ krsq-crf) mulps xmm5, [esp + nb212_two] subps xmm7, xmm5 ;# xmm7=rinv-2*krsq mulps xmm7, [esp + nb212_qqOO] ;# xmm7 = coul part of fscal movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=rinvtwelve mulps xmm1, [esp + nb212_c6] mulps xmm2, [esp + nb212_c12] movaps xmm3, xmm2 subps xmm3, xmm1 ;# xmm3=Vvdw12-Vvdw6 addps xmm3, [esp + nb212_Vvdwtot] mulps xmm1, [esp + nb212_six] mulps xmm2, [esp + nb212_twelve] movaps [esp + nb212_Vvdwtot], xmm3 subps xmm2, xmm1 addps xmm2, xmm7 addps xmm6, [esp + nb212_vctot] ;# local vctot summation variable mulps xmm0, xmm2 movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb212_dxOO] mulps xmm1, [esp + nb212_dyOO] mulps xmm2, [esp + nb212_dzOO] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb212_fixO] addps xmm1, [esp + nb212_fiyO] addps xmm2, [esp + nb212_fizO] movaps [esp + nb212_fjxO], xmm3 movaps [esp + nb212_fjyO], xmm4 movaps [esp + nb212_fjzO], xmm5 movaps [esp + nb212_fixO], xmm0 movaps [esp + nb212_fiyO], xmm1 movaps [esp + nb212_fizO], xmm2 ;# O-H1 interaction movaps xmm0, [esp + nb212_rinvOH1] movaps xmm7, xmm0 ;# xmm7=rinv movaps xmm5, [esp + nb212_krf] movaps xmm1, xmm0 mulps xmm5, [esp + nb212_rsqOH1] ;# xmm5=krsq movaps xmm4, xmm5 addps xmm4, xmm7 ;# xmm4=rinv+ krsq mulps xmm0, xmm0 subps xmm4, [esp + nb212_crf] mulps xmm4, [esp + nb212_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) mulps xmm5, [esp + nb212_two] subps xmm7, xmm5 ;# xmm7=rinv-2*krsq mulps xmm7, [esp + nb212_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 + nb212_dxOH1] mulps xmm1, [esp + nb212_dyOH1] mulps xmm2, [esp + nb212_dzOH1] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb212_fixO] addps xmm1, [esp + nb212_fiyO] addps xmm2, [esp + nb212_fizO] movaps [esp + nb212_fjxH1], xmm3 movaps [esp + nb212_fjyH1], xmm4 movaps [esp + nb212_fjzH1], xmm5 movaps [esp + nb212_fixO], xmm0 movaps [esp + nb212_fiyO], xmm1 movaps [esp + nb212_fizO], xmm2 ;# O-H2 interaction movaps xmm0, [esp + nb212_rinvOH2] movaps xmm7, xmm0 ;# xmm7=rinv movaps xmm5, [esp + nb212_krf] movaps xmm1, xmm0 mulps xmm5, [esp + nb212_rsqOH2] ;# xmm5=krsq movaps xmm4, xmm5 addps xmm4, xmm7 ;# xmm4=r inv+ krsq mulps xmm0, xmm0 subps xmm4, [esp + nb212_crf] mulps xmm4, [esp + nb212_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) mulps xmm5, [esp + nb212_two] subps xmm7, xmm5 ;# xmm7=rinv-2*krsq mulps xmm7, [esp + nb212_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 + nb212_dxOH2] mulps xmm1, [esp + nb212_dyOH2] mulps xmm2, [esp + nb212_dzOH2] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb212_fixO] addps xmm1, [esp + nb212_fiyO] addps xmm2, [esp + nb212_fizO] movaps [esp + nb212_fjxH2], xmm3 movaps [esp + nb212_fjyH2], xmm4 movaps [esp + nb212_fjzH2], xmm5 movaps [esp + nb212_fixO], xmm0 movaps [esp + nb212_fiyO], xmm1 movaps [esp + nb212_fizO], xmm2 ;# H1-O interaction movaps xmm0, [esp + nb212_rinvH1O] movaps xmm7, xmm0 ;# xmm7=rinv movaps xmm5, [esp + nb212_krf] movaps xmm1, xmm0
⌨️ 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?