nb_kernel112_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,169 行 · 第 1/5 页
S
2,169 行
movapd xmm0, [esp + nb112_rinvH1O] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb112_qqOH] mulpd xmm0, xmm1 ;# fsH1O addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb112_fjxO] movapd xmm4, [esp + nb112_fjyO] movapd xmm5, [esp + nb112_fjzO] mulpd xmm0, [esp + nb112_dxH1O] mulpd xmm1, [esp + nb112_dyH1O] mulpd xmm2, [esp + nb112_dzH1O] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb112_fixH1] addpd xmm1, [esp + nb112_fiyH1] addpd xmm2, [esp + nb112_fizH1] movapd [esp + nb112_fjxO], xmm3 movapd [esp + nb112_fjyO], xmm4 movapd [esp + nb112_fjzO], xmm5 movapd [esp + nb112_fixH1], xmm0 movapd [esp + nb112_fiyH1], xmm1 movapd [esp + nb112_fizH1], xmm2 ;# H1-H1 interaction movapd xmm0, [esp + nb112_rinvH1H1] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb112_qqHH] mulpd xmm0, xmm1 ;# fsH1H1 addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb112_fjxH1] movapd xmm4, [esp + nb112_fjyH1] movapd xmm5, [esp + nb112_fjzH1] mulpd xmm0, [esp + nb112_dxH1H1] mulpd xmm1, [esp + nb112_dyH1H1] mulpd xmm2, [esp + nb112_dzH1H1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb112_fixH1] addpd xmm1, [esp + nb112_fiyH1] addpd xmm2, [esp + nb112_fizH1] movapd [esp + nb112_fjxH1], xmm3 movapd [esp + nb112_fjyH1], xmm4 movapd [esp + nb112_fjzH1], xmm5 movapd [esp + nb112_fixH1], xmm0 movapd [esp + nb112_fiyH1], xmm1 movapd [esp + nb112_fizH1], xmm2 ;# H1-H2 interaction movapd xmm0, [esp + nb112_rinvH1H2] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb112_qqHH] mulpd xmm0, xmm1 ;# fsOH2 addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb112_fjxH2] movapd xmm4, [esp + nb112_fjyH2] movapd xmm5, [esp + nb112_fjzH2] mulpd xmm0, [esp + nb112_dxH1H2] mulpd xmm1, [esp + nb112_dyH1H2] mulpd xmm2, [esp + nb112_dzH1H2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb112_fixH1] addpd xmm1, [esp + nb112_fiyH1] addpd xmm2, [esp + nb112_fizH1] movapd [esp + nb112_fjxH2], xmm3 movapd [esp + nb112_fjyH2], xmm4 movapd [esp + nb112_fjzH2], xmm5 movapd [esp + nb112_fixH1], xmm0 movapd [esp + nb112_fiyH1], xmm1 movapd [esp + nb112_fizH1], xmm2 ;# H2-O interaction movapd xmm0, [esp + nb112_rinvH2O] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb112_qqOH] mulpd xmm0, xmm1 ;# fsH2O addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb112_fjxO] movapd xmm4, [esp + nb112_fjyO] movapd xmm5, [esp + nb112_fjzO] mulpd xmm0, [esp + nb112_dxH2O] mulpd xmm1, [esp + nb112_dyH2O] mulpd xmm2, [esp + nb112_dzH2O] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb112_fixH2] addpd xmm1, [esp + nb112_fiyH2] addpd xmm2, [esp + nb112_fizH2] movapd [esp + nb112_fjxO], xmm3 movapd [esp + nb112_fjyO], xmm4 movapd [esp + nb112_fjzO], xmm5 movapd [esp + nb112_fixH2], xmm0 movapd [esp + nb112_fiyH2], xmm1 movapd [esp + nb112_fizH2], xmm2 ;# H2-H1 interaction movapd xmm0, [esp + nb112_rinvH2H1] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb112_qqHH] mulpd xmm0, xmm1 ;# fsH2H1 addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb112_fjxH1] movapd xmm4, [esp + nb112_fjyH1] movapd xmm5, [esp + nb112_fjzH1] mulpd xmm0, [esp + nb112_dxH2H1] mulpd xmm1, [esp + nb112_dyH2H1] mulpd xmm2, [esp + nb112_dzH2H1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb112_fixH2] addpd xmm1, [esp + nb112_fiyH2] addpd xmm2, [esp + nb112_fizH2] movapd [esp + nb112_fjxH1], xmm3 movapd [esp + nb112_fjyH1], xmm4 movapd [esp + nb112_fjzH1], xmm5 movapd [esp + nb112_fixH2], xmm0 movapd [esp + nb112_fiyH2], xmm1 movapd [esp + nb112_fizH2], xmm2 ;# H2-H2 interaction movapd xmm0, [esp + nb112_rinvH2H2] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb112_qqHH] mulpd xmm0, xmm1 ;# fsH2H2 addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd [esp + nb112_vctot], xmm7 movapd xmm2, xmm0 movapd xmm3, [esp + nb112_fjxH2] movapd xmm4, [esp + nb112_fjyH2] movapd xmm5, [esp + nb112_fjzH2] mulpd xmm0, [esp + nb112_dxH2H2] mulpd xmm1, [esp + nb112_dyH2H2] mulpd xmm2, [esp + nb112_dzH2H2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb112_fixH2] addpd xmm1, [esp + nb112_fiyH2] addpd xmm2, [esp + nb112_fizH2] movapd [esp + nb112_fjxH2], xmm3 movapd [esp + nb112_fjyH2], xmm4 movapd [esp + nb112_fjzH2], xmm5 movapd [esp + nb112_fixH2], xmm0 movapd [esp + nb112_fiyH2], xmm1 movapd [esp + nb112_fizH2], xmm2 mov edi, [ebp + nb112_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 + nb112_fjxO] addpd xmm1, [esp + nb112_fjyO] addpd xmm2, [esp + nb112_fjzO] addpd xmm3, [esp + nb112_fjxH1] addpd xmm4, [esp + nb112_fjyH1] addpd xmm5, [esp + nb112_fjzH1] addpd xmm6, [esp + nb112_fjxH2] addpd xmm7, [esp + nb112_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 + nb112_fjzH2] movlpd [edi + eax*8 + 64], xmm0 movhpd [edi + ebx*8 + 64], xmm0 ;# should we do one more iteration? sub dword ptr [esp + nb112_innerk], 2 jl .nb112_checksingle jmp .nb112_unroll_loop.nb112_checksingle: mov edx, [esp + nb112_innerk] and edx, 1 jnz .nb112_dosingle jmp .nb112_updateouterdata.nb112_dosingle: mov edx, [esp + nb112_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] add dword ptr [esp + nb112_innerjjnr], 4 mov esi, [ebp + nb112_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 + nb112_jxO], xmm2 movapd [esp + nb112_jyO], xmm3 movapd [esp + nb112_jzO], xmm4 movapd [esp + nb112_jxH1], xmm5 movapd [esp + nb112_jyH1], xmm6 movapd [esp + nb112_jzH1], xmm7 movlpd xmm2, [esi + eax*8 + 48] movlpd xmm3, [esi + eax*8 + 56] movlpd xmm4, [esi + eax*8 + 64] movapd [esp + nb112_jxH2], xmm2 movapd [esp + nb112_jyH2], xmm3 movapd [esp + nb112_jzH2], xmm4 movapd xmm0, [esp + nb112_ixO] movapd xmm1, [esp + nb112_iyO] movapd xmm2, [esp + nb112_izO] movapd xmm3, [esp + nb112_ixO] movapd xmm4, [esp + nb112_iyO] movapd xmm5, [esp + nb112_izO] subsd xmm0, [esp + nb112_jxO] subsd xmm1, [esp + nb112_jyO] subsd xmm2, [esp + nb112_jzO] subsd xmm3, [esp + nb112_jxH1] subsd xmm4, [esp + nb112_jyH1] subsd xmm5, [esp + nb112_jzH1] movapd [esp + nb112_dxOO], xmm0 movapd [esp + nb112_dyOO], xmm1 movapd [esp + nb112_dzOO], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb112_dxOH1], xmm3 movapd [esp + nb112_dyOH1], xmm4 movapd [esp + nb112_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 + nb112_rsqOO], xmm0 movapd [esp + nb112_rsqOH1], xmm3 movapd xmm0, [esp + nb112_ixO] movapd xmm1, [esp + nb112_iyO] movapd xmm2, [esp + nb112_izO] movapd xmm3, [esp + nb112_ixH1] movapd xmm4, [esp + nb112_iyH1] movapd xmm5, [esp + nb112_izH1] subsd xmm0, [esp + nb112_jxH2] subsd xmm1, [esp + nb112_jyH2] subsd xmm2, [esp + nb112_jzH2] subsd xmm3, [esp + nb112_jxO] subsd xmm4, [esp + nb112_jyO] subsd xmm5, [esp + nb112_jzO] movapd [esp + nb112_dxOH2], xmm0 movapd [esp + nb112_dyOH2], xmm1 movapd [esp + nb112_dzOH2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb112_dxH1O], xmm3 movapd [esp + nb112_dyH1O], xmm4 movapd [esp + nb112_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 + nb112_rsqOH2], xmm0 movapd [esp + nb112_rsqH1O], xmm3 movapd xmm0, [esp + nb112_ixH1] movapd xmm1, [esp + nb112_iyH1] movapd xmm2, [esp + nb112_izH1] movapd xmm3, [esp + nb112_ixH1] movapd xmm4, [esp + nb112_iyH1] movapd xmm5, [esp + nb112_izH1] subsd xmm0, [esp + nb112_jxH1] subsd xmm1, [esp + nb112_jyH1] subsd xmm2, [esp + nb112_jzH1] subsd xmm3, [esp + nb112_jxH2] subsd xmm4, [esp + nb112_jyH2] subsd xmm5, [esp + nb112_jzH2] movapd [esp + nb112_dxH1H1], xmm0 movapd [esp + nb112_dyH1H1], xmm1 movapd [esp + nb112_dzH1H1], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb112_dxH1H2], xmm3 movapd [esp + nb112_dyH1H2], xmm4 movapd [esp + nb112_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 + nb112_rsqH1H1], xmm0 movapd [esp + nb112_rsqH1H2], xmm3 movapd xmm0, [esp + nb112_ixH2] movapd xmm1, [esp + nb112_iyH2] movapd xmm2, [esp + nb112_izH2] movapd xmm3, [esp + nb112_ixH2] movapd xmm4, [esp + nb112_iyH2] movapd xmm5, [esp + nb112_izH2] subsd xmm0, [esp + nb112_jxO] subsd xmm1, [esp + nb112_jyO] subsd xmm2, [esp + nb112_jzO] subsd xmm3, [esp + nb112_jxH1] subsd xmm4, [esp + nb112_jyH1] subsd xmm5, [esp + nb112_jzH1] movapd [esp + nb112_dxH2O], xmm0 movapd [esp + nb112_dyH2O], xmm1 movapd [esp + nb112_dzH2O], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb112_dxH2H1], xmm3 movapd [esp + nb112_dyH2H1], xmm4 movapd [esp + nb112_dzH2H1], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm4, xmm3 addsd xmm4, xmm5 movapd [esp + nb112_rsqH2O], xmm0 movapd [esp + nb112_rsqH2H1], xmm4 movapd xmm0, [esp + nb112_ixH2] movapd xmm1, [esp + nb112_iyH2] movapd xmm2, [esp + nb112_izH2] subsd xmm0, [esp + nb112_jxH2] subsd xmm1, [esp + nb112_jyH2] subsd xmm2, [esp + nb112_jzH2] movapd [esp + nb112_dxH2H2], xmm0 movapd [esp + nb112_dyH2H2], xmm1 movapd [esp + nb112_dzH2H2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 addsd xmm0, xmm1 addsd xmm0, xmm2 movapd [esp + nb112_rsqH2H2], xmm0 ;# start doing invsqrt use rsq values in xmm0, xmm4 cvtsd2ss xmm1, xmm0 cvtsd2ss xmm5, xmm4 rsqrtss xmm1, xmm1 rsqrtss xmm5, xmm5 cvtss2sd xmm1, xmm1 cvtss2sd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulsd xmm1, xmm1 ;# luA*luA mulsd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb112_three] mulsd xmm1, xmm0 ;# rsqA*luA*luA mulsd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subsd xmm3, xmm1 ;# 3-rsqA*luA*luA subsd xmm7, xmm5 ;# 3-rsqB*luB*luB mulsd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulsd xmm3, [esp + nb112_half] ;# iter1 mulsd xmm7, [esp + nb112_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulsd xmm3, xmm3 ;# luA*luA mulsd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb112_three] mulsd xmm3, xmm0 ;# rsqA*luA*luA mulsd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subsd xmm1, xmm3 ;# 3-rsqA*luA*luA subsd xmm5, xmm7 ;# 3-rsqB*luB*luB mulsd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulsd xmm1, [esp + nb112_half] ;# rinv mulsd xmm5, [esp + nb112_half] ;# rinv movapd [esp + nb112_rinvH2H2], xmm1 movapd [esp + nb112_rinvH2H1], xmm5
⌨️ 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?