nb_kernel202_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,125 行 · 第 1/5 页
S
2,125 行
subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb202_fixO] addpd xmm1, [esp + nb202_fiyO] addpd xmm2, [esp + nb202_fizO] movapd [esp + nb202_fjxH2], xmm3 movapd [esp + nb202_fjyH2], xmm4 movapd [esp + nb202_fjzH2], xmm5 movapd [esp + nb202_fixO], xmm0 movapd [esp + nb202_fiyO], xmm1 movapd [esp + nb202_fizO], xmm2 ;# H1-O interaction movapd xmm0, [esp + nb202_rinvH1O] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb202_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb202_rsqH1O] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=rinv+ krsq mulpd xmm0, xmm0 subpd xmm4, [esp + nb202_crf] mulpd xmm4, [esp + nb202_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb202_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb202_qqOH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb202_fjxO] movapd xmm4, [esp + nb202_fjyO] movapd xmm5, [esp + nb202_fjzO] mulpd xmm0, [esp + nb202_dxH1O] mulpd xmm1, [esp + nb202_dyH1O] mulpd xmm2, [esp + nb202_dzH1O] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb202_fixH1] addpd xmm1, [esp + nb202_fiyH1] addpd xmm2, [esp + nb202_fizH1] movapd [esp + nb202_fjxO], xmm3 movapd [esp + nb202_fjyO], xmm4 movapd [esp + nb202_fjzO], xmm5 movapd [esp + nb202_fixH1], xmm0 movapd [esp + nb202_fiyH1], xmm1 movapd [esp + nb202_fizH1], xmm2 ;# H1-H1 interaction movapd xmm0, [esp + nb202_rinvH1H1] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb202_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb202_rsqH1H1] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb202_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb202_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb202_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb202_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb202_fjxH1] movapd xmm4, [esp + nb202_fjyH1] movapd xmm5, [esp + nb202_fjzH1] mulpd xmm0, [esp + nb202_dxH1H1] mulpd xmm1, [esp + nb202_dyH1H1] mulpd xmm2, [esp + nb202_dzH1H1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb202_fixH1] addpd xmm1, [esp + nb202_fiyH1] addpd xmm2, [esp + nb202_fizH1] movapd [esp + nb202_fjxH1], xmm3 movapd [esp + nb202_fjyH1], xmm4 movapd [esp + nb202_fjzH1], xmm5 movapd [esp + nb202_fixH1], xmm0 movapd [esp + nb202_fiyH1], xmm1 movapd [esp + nb202_fizH1], xmm2 ;# H1-H2 interaction movapd xmm0, [esp + nb202_rinvH1H2] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb202_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb202_rsqH1H2] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq mulpd xmm0, xmm0 subpd xmm4, [esp + nb202_crf] mulpd xmm4, [esp + nb202_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb202_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb202_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb202_fjxH2] movapd xmm4, [esp + nb202_fjyH2] movapd xmm5, [esp + nb202_fjzH2] mulpd xmm0, [esp + nb202_dxH1H2] mulpd xmm1, [esp + nb202_dyH1H2] mulpd xmm2, [esp + nb202_dzH1H2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb202_fixH1] addpd xmm1, [esp + nb202_fiyH1] addpd xmm2, [esp + nb202_fizH1] movapd [esp + nb202_fjxH2], xmm3 movapd [esp + nb202_fjyH2], xmm4 movapd [esp + nb202_fjzH2], xmm5 movapd [esp + nb202_fixH1], xmm0 movapd [esp + nb202_fiyH1], xmm1 movapd [esp + nb202_fizH1], xmm2 ;# H2-O interaction movapd xmm0, [esp + nb202_rinvH2O] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb202_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb202_rsqH2O] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb202_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb202_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb202_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb202_qqOH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb202_fjxO] movapd xmm4, [esp + nb202_fjyO] movapd xmm5, [esp + nb202_fjzO] mulpd xmm0, [esp + nb202_dxH2O] mulpd xmm1, [esp + nb202_dyH2O] mulpd xmm2, [esp + nb202_dzH2O] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb202_fixH2] addpd xmm1, [esp + nb202_fiyH2] addpd xmm2, [esp + nb202_fizH2] movapd [esp + nb202_fjxO], xmm3 movapd [esp + nb202_fjyO], xmm4 movapd [esp + nb202_fjzO], xmm5 movapd [esp + nb202_fixH2], xmm0 movapd [esp + nb202_fiyH2], xmm1 movapd [esp + nb202_fizH2], xmm2 ;# H2-H1 interaction movapd xmm0, [esp + nb202_rinvH2H1] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb202_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb202_rsqH2H1] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb202_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb202_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb202_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb202_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb202_fjxH1] movapd xmm4, [esp + nb202_fjyH1] movapd xmm5, [esp + nb202_fjzH1] mulpd xmm0, [esp + nb202_dxH2H1] mulpd xmm1, [esp + nb202_dyH2H1] mulpd xmm2, [esp + nb202_dzH2H1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb202_fixH2] addpd xmm1, [esp + nb202_fiyH2] addpd xmm2, [esp + nb202_fizH2] movapd [esp + nb202_fjxH1], xmm3 movapd [esp + nb202_fjyH1], xmm4 movapd [esp + nb202_fjzH1], xmm5 movapd [esp + nb202_fixH2], xmm0 movapd [esp + nb202_fiyH2], xmm1 movapd [esp + nb202_fizH2], xmm2 ;# H2-H2 interaction movapd xmm0, [esp + nb202_rinvH2H2] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb202_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb202_rsqH2H2] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb202_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb202_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb202_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb202_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm1, xmm0 movapd [esp + nb202_vctot], xmm6 movapd xmm2, xmm0 movapd xmm3, [esp + nb202_fjxH2] movapd xmm4, [esp + nb202_fjyH2] movapd xmm5, [esp + nb202_fjzH2] mulpd xmm0, [esp + nb202_dxH2H2] mulpd xmm1, [esp + nb202_dyH2H2] mulpd xmm2, [esp + nb202_dzH2H2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb202_fixH2] addpd xmm1, [esp + nb202_fiyH2] addpd xmm2, [esp + nb202_fizH2] movapd [esp + nb202_fjxH2], xmm3 movapd [esp + nb202_fjyH2], xmm4 movapd [esp + nb202_fjzH2], xmm5 movapd [esp + nb202_fixH2], xmm0 movapd [esp + nb202_fiyH2], xmm1 movapd [esp + nb202_fizH2], xmm2 mov edi, [ebp + nb202_faction] ;# Did all interactions - now update j forces movlpd xmm0, [edi + eax*8] movlpd xmm1, [edi + eax*8 + 8] movlpd xmm2, [edi + eax*8 + 16] movlpd xmm3, [edi + eax*8 + 24] movlpd xmm4, [edi + eax*8 + 32] movlpd xmm5, [edi + eax*8 + 40] movlpd xmm6, [edi + eax*8 + 48] movlpd xmm7, [edi + eax*8 + 56] movhpd xmm0, [edi + ebx*8] movhpd xmm1, [edi + ebx*8 + 8] movhpd xmm2, [edi + ebx*8 + 16] movhpd xmm3, [edi + ebx*8 + 24] movhpd xmm4, [edi + ebx*8 + 32] movhpd xmm5, [edi + ebx*8 + 40] movhpd xmm6, [edi + ebx*8 + 48] movhpd xmm7, [edi + ebx*8 + 56] addpd xmm0, [esp + nb202_fjxO] addpd xmm1, [esp + nb202_fjyO] addpd xmm2, [esp + nb202_fjzO] addpd xmm3, [esp + nb202_fjxH1] addpd xmm4, [esp + nb202_fjyH1] addpd xmm5, [esp + nb202_fjzH1] addpd xmm6, [esp + nb202_fjxH2] addpd xmm7, [esp + nb202_fjyH2] movlpd [edi + eax*8], xmm0 movlpd [edi + eax*8 + 8], xmm1 movlpd [edi + eax*8 + 16], xmm2 movlpd [edi + eax*8 + 24], xmm3 movlpd [edi + eax*8 + 32], xmm4 movlpd [edi + eax*8 + 40], xmm5 movlpd [edi + eax*8 + 48], xmm6 movlpd [edi + eax*8 + 56], xmm7 movhpd [edi + ebx*8], xmm0 movhpd [edi + ebx*8 + 8], xmm1 movhpd [edi + ebx*8 + 16], xmm2 movhpd [edi + ebx*8 + 24], xmm3 movhpd [edi + ebx*8 + 32], xmm4 movhpd [edi + ebx*8 + 40], xmm5 movhpd [edi + ebx*8 + 48], xmm6 movhpd [edi + ebx*8 + 56], xmm7 movlpd xmm0, [edi + eax*8 + 64] movhpd xmm0, [edi + ebx*8 + 64] addpd xmm0, [esp + nb202_fjzH2] movlpd [edi + eax*8 + 64], xmm0 movhpd [edi + ebx*8 + 64], xmm0 ;# should we do one more iteration? sub dword ptr [esp + nb202_innerk], 2 jl .nb202_checksingle jmp .nb202_unroll_loop.nb202_checksingle: mov edx, [esp + nb202_innerk] and edx, 1 jnz .nb202_dosingle jmp .nb202_updateouterdata.nb202_dosingle: mov edx, [esp + nb202_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] mov esi, [ebp + nb202_pos] lea eax, [eax + eax*2] ;# fetch j coordinates movlpd xmm2, [esi + eax*8] movlpd xmm3, [esi + eax*8 + 8] movlpd xmm4, [esi + eax*8 + 16] movlpd xmm5, [esi + eax*8 + 24] movlpd xmm6, [esi + eax*8 + 32] movlpd xmm7, [esi + eax*8 + 40] movapd [esp + nb202_jxO], xmm2 movapd [esp + nb202_jyO], xmm3 movapd [esp + nb202_jzO], xmm4 movapd [esp + nb202_jxH1], xmm5 movapd [esp + nb202_jyH1], xmm6 movapd [esp + nb202_jzH1], xmm7 movlpd xmm2, [esi + eax*8 + 48] movlpd xmm3, [esi + eax*8 + 56] movlpd xmm4, [esi + eax*8 + 64] movapd [esp + nb202_jxH2], xmm2 movapd [esp + nb202_jyH2], xmm3 movapd [esp + nb202_jzH2], xmm4 movapd xmm0, [esp + nb202_ixO] movapd xmm1, [esp + nb202_iyO] movapd xmm2, [esp + nb202_izO] movapd xmm3, [esp + nb202_ixO] movapd xmm4, [esp + nb202_iyO] movapd xmm5, [esp + nb202_izO] subsd xmm0, [esp + nb202_jxO] subsd xmm1, [esp + nb202_jyO] subsd xmm2, [esp + nb202_jzO] subsd xmm3, [esp + nb202_jxH1] subsd xmm4, [esp + nb202_jyH1] subsd xmm5, [esp + nb202_jzH1] movapd [esp + nb202_dxOO], xmm0 movapd [esp + nb202_dyOO], xmm1 movapd [esp + nb202_dzOO], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb202_dxOH1], xmm3 movapd [esp + nb202_dyOH1], xmm4 movapd [esp + nb202_dzOH1], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd xmm3, xmm5 movapd [esp + nb202_rsqOO], xmm0 movapd [esp + nb202_rsqOH1], xmm3 movapd xmm0, [esp + nb202_ixO] movapd xmm1, [esp + nb202_iyO] movapd xmm2, [esp + nb202_izO] movapd xmm3, [esp + nb202_ixH1] movapd xmm4, [esp + nb202_iyH1] movapd xmm5, [esp + nb202_izH1] subsd xmm0, [esp + nb202_jxH2] subsd xmm1, [esp + nb202_jyH2] subsd xmm2, [esp + nb202_jzH2] subsd xmm3, [esp + nb202_jxO] subsd xmm4, [esp + nb202_jyO] subsd xmm5, [esp + nb202_jzO] movapd [esp + nb202_dxOH2], xmm0 movapd [esp + nb202_dyOH2], xmm1 movapd [esp + nb202_dzOH2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb202_dxH1O], xmm3 movapd [esp + nb202_dyH1O], xmm4 movapd [esp + nb202_dzH1O], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd xmm3, xmm5 movapd [esp + nb202_rsqOH2], xmm0 movapd [esp + nb202_rsqH1O], xmm3 movapd xmm0, [esp + nb202_ixH1] movapd xmm1, [esp + nb202_iyH1] movapd xmm2, [esp + nb202_izH1] movapd xmm3, [esp + nb202_ixH1] movapd xmm4, [esp + nb202_iyH1] movapd xmm5, [esp + nb202_izH1] subsd xmm0, [esp + nb202_jxH1] subsd xmm1, [esp + nb202_jyH1] subsd xmm2, [esp + nb202_jzH1] subsd xmm3, [esp + nb202_jxH2] subsd xmm4, [esp + nb202_jyH2] subsd xmm5, [esp + nb202_jzH2] movapd [esp + nb202_dxH1H1], xmm0 movapd [esp + nb202_dyH1H1], xmm1 movapd [esp + nb202_dzH1H1], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb202_dxH1H2], xmm3 movapd [esp + nb202_dyH1H2], xmm4 movapd [esp + nb202_dzH1H2], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd xmm3, xmm5 movapd [esp + nb202_rsqH1H1], xmm0 movapd [esp + nb202_rsqH1H2], xmm3 movapd xmm0, [esp + nb202_ixH2]
⌨️ 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?