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_rsqOO] movapd xmm4, [esp + nb112_rsqOH1] 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 of mulsd xmm7, [esp + nb112_half] ;# iter1 of 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_rinvOO], xmm1 movapd [esp + nb112_rinvOH1], xmm5 movapd xmm0, [esp + nb112_rsqOH2] movapd xmm4, [esp + nb112_rsqH1O] 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_rinvOH2], xmm1 movapd [esp + nb112_rinvH1O], xmm5 movapd xmm0, [esp + nb112_rsqH1H1] movapd xmm4, [esp + nb112_rsqH1H2] 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] ;# iter1a mulsd xmm7, [esp + nb112_half] ;# iter1b 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_rinvH1H1], xmm1 movapd [esp + nb112_rinvH1H2], xmm5 movapd xmm0, [esp + nb112_rsqH2O] cvtsd2ss xmm1, xmm0 rsqrtss xmm1, xmm1 cvtss2sd xmm1, xmm1 movapd xmm2, xmm1 ;# copy of luA mulsd xmm1, xmm1 ;# luA*luA movapd xmm3, [esp + nb112_three] mulsd xmm1, xmm0 ;# rsqA*luA*luA subsd xmm3, xmm1 ;# 3-rsqA*luA*luA mulsd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm3, [esp + nb112_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA mulsd xmm3, xmm3 ;# luA*luA movapd xmm1, [esp + nb112_three] mulsd xmm3, xmm0 ;# rsqA*luA*luA subsd xmm1, xmm3 ;# 3-rsqA*luA*luA mulsd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulsd xmm1, [esp + nb112_half] ;# rinv movapd [esp + nb112_rinvH2O], xmm1 ;# start with OO interaction movapd xmm0, [esp + nb112_rinvOO] movapd xmm7, xmm0 mulsd xmm0, xmm0 movapd xmm1, xmm0 mulsd xmm1, xmm0 mulsd xmm1, xmm0 ;# xmm1=rinvsix mulsd xmm7, [esp + nb112_qqOO] movapd xmm2, xmm1 mulsd xmm2, xmm2 ;# xmm2=rinvtwelve mulsd xmm1, [esp + nb112_c6] mulsd xmm2, [esp + nb112_c12] movapd xmm3, xmm2 subsd xmm3, xmm1 ;# xmm3=Vvdw12-Vvdw6 addsd xmm3, [esp + nb112_Vvdwtot] mulsd xmm1, [esp + nb112_six] mulsd xmm2, [esp + nb112_twelve] movlpd [esp + nb112_Vvdwtot], xmm3 subsd xmm2, xmm1 addsd xmm2, xmm7 addsd xmm7, [esp + nb112_vctot] mulsd xmm0, xmm2 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb112_dxOO] mulsd xmm1, [esp + nb112_dyOO] mulsd xmm2, [esp + nb112_dzOO] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb112_fixO] addsd xmm1, [esp + nb112_fiyO] addsd xmm2, [esp + nb112_fizO] movlpd [esp + nb112_fjxO], xmm3 movlpd [esp + nb112_fjyO], xmm4 movlpd [esp + nb112_fjzO], xmm5 movlpd [esp + nb112_fixO], xmm0 movlpd [esp + nb112_fiyO], xmm1 movlpd [esp + nb112_fizO], xmm2 ;# O-H1 interaction movapd xmm0, [esp + nb112_rinvOH1] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb112_qqOH] mulsd xmm0, xmm1 ;# fsOH1 addsd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb112_dxOH1] mulsd xmm1, [esp + nb112_dyOH1] mulsd xmm2, [esp + nb112_dzOH1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb112_fixO] addsd xmm1, [esp + nb112_fiyO] addsd xmm2, [esp + nb112_fizO] movlpd [esp + nb112_fjxH1], xmm3 movlpd [esp + nb112_fjyH1], xmm4 movlpd [esp + nb112_fjzH1], xmm5 movlpd [esp + nb112_fixO], xmm0 movlpd [esp + nb112_fiyO], xmm1 movlpd [esp + nb112_fizO], xmm2 ;# O-H2 interaction movapd xmm0, [esp + nb112_rinvOH2] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb112_qqOH] mulsd xmm0, xmm1 ;# fsOH2 addsd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb112_dxOH2] mulsd xmm1, [esp + nb112_dyOH2] mulsd xmm2, [esp + nb112_dzOH2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb112_fixO] addsd xmm1, [esp + nb112_fiyO] addsd xmm2, [esp + nb112_fizO] movlpd [esp + nb112_fjxH2], xmm3 movlpd [esp + nb112_fjyH2], xmm4 movlpd [esp + nb112_fjzH2], xmm5 movlpd [esp + nb112_fixO], xmm0 movlpd [esp + nb112_fiyO], xmm1 movlpd [esp + nb112_fizO], xmm2 ;# H1-O interaction movapd xmm0, [esp + nb112_rinvH1O] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb112_qqOH] mulsd xmm0, xmm1 ;# fsH1O addsd 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] mulsd xmm0, [esp + nb112_dxH1O] mulsd xmm1, [esp + nb112_dyH1O] mulsd xmm2, [esp + nb112_dzH1O] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb112_fixH1] addsd xmm1, [esp + nb112_fiyH1] addsd xmm2, [esp + nb112_fizH1] movlpd [esp + nb112_fjxO], xmm3 movlpd [esp + nb112_fjyO], xmm4 movlpd [esp + nb112_fjzO], xmm5 movlpd [esp + nb112_fixH1], xmm0 movlpd [esp + nb112_fiyH1], xmm1 movlpd [esp + nb112_fizH1], xmm2 ;# H1-H1 interaction movapd xmm0, [esp + nb112_rinvH1H1] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb112_qqHH] mulsd xmm0, xmm1 ;# fsH1H1 addsd 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] mulsd xmm0, [esp + nb112_dxH1H1] mulsd xmm1, [esp + nb112_dyH1H1] mulsd xmm2, [esp + nb112_dzH1H1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb112_fixH1] addsd xmm1, [esp + nb112_fiyH1] addsd xmm2, [esp + nb112_fizH1] movlpd [esp + nb112_fjxH1], xmm3 movlpd [esp + nb112_fjyH1], xmm4 movlpd [esp + nb112_fjzH1], xmm5 movlpd [esp + nb112_fixH1], xmm0 movlpd [esp + nb112_fiyH1], xmm1 movlpd [esp + nb112_fizH1], xmm2 ;# H1-H2 interaction movapd xmm0, [esp + nb112_rinvH1H2] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb112_qqHH] mulsd xmm0, xmm1 ;# fsOH2 addsd 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] mulsd xmm0, [esp + nb112_dxH1H2] mulsd xmm1, [esp + nb112_dyH1H2] mulsd xmm2, [esp + nb112_dzH1H2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb112_fixH1] addsd xmm1, [esp + nb112_fiyH1] addsd xmm2, [esp + nb112_fizH1] movlpd [esp + nb112_fjxH2], xmm3 movlpd [esp + nb112_fjyH2], xmm4 movlpd [esp + nb112_fjzH2], xmm5 movlpd [esp + nb112_fixH1], xmm0 movlpd [esp + nb112_fiyH1], xmm1 movlpd [esp + nb112_fizH1], xmm2 ;# H2-O interaction movapd xmm0, [esp + nb112_rinvH2O] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb112_qqOH] mulsd xmm0, xmm1 ;# fsH2O addsd 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] mulsd xmm0, [esp + nb112_dxH2O] mulsd xmm1, [esp + nb112_dyH2O] mulsd xmm2, [esp + nb112_dzH2O] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb112_fixH2] addsd xmm1, [esp + nb112_fiyH2] addsd xmm2, [esp + nb112_fizH2] movlpd [esp + nb112_fjxO], xmm3 movlpd [esp + nb112_fjyO], xmm4 movlpd [esp + nb112_fjzO], xmm5 movlpd [esp + nb112_fixH2], xmm0 movlpd [esp + nb112_fiyH2], xmm1 movlpd [esp + nb112_fizH2], xmm2 ;# H2-H1 interaction movapd xmm0, [esp + nb112_rinvH2H1] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb112_qqHH] mulsd xmm0, xmm1 ;# fsH2H1 addsd 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] mulsd xmm0, [esp + nb112_dxH2H1] mulsd xmm1, [esp + nb112_dyH2H1] mulsd xmm2, [esp + nb112_dzH2H1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb112_fixH2] addsd xmm1, [esp + nb112_fiyH2] addsd xmm2, [esp + nb112_fizH2] movlpd [esp + nb112_fjxH1], xmm3 movlpd [esp + nb112_fjyH1], xmm4 movlpd [esp + nb112_fjzH1], xmm5 movlpd [esp + nb112_fixH2], xmm0 movlpd [esp + nb112_fiyH2], xmm1 movlpd [esp + nb112_fizH2], xmm2 ;# H2-H2 interaction movapd xmm0, [esp + nb112_rinvH2H2] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb112_qqHH] mulsd xmm0, xmm1 ;# fsH2H2 addsd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movlpd [esp + nb112_vctot], xmm7 movapd xmm2, xmm0 movlpd xmm3, [esp + nb112_fjxH2] movlpd xmm4, [esp + nb112_fjyH2] movlpd xmm5, [esp + nb112_fjzH2] mulsd xmm0, [esp + nb112_dxH2H2] mulsd xmm1, [esp + nb112_dyH2H2] mulsd xmm2, [esp + nb112_dzH2H2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb112_fixH2] addsd xmm1, [esp + nb112_fiyH2] addsd xmm2, [esp + nb112_fizH2] movlpd [esp + nb112_fjxH2], xmm3 movlpd [esp + nb112_fjyH2], xmm4 movlpd [esp + nb112_fjzH2], xmm5 movlpd [esp + nb112_fixH2], xmm0 movlpd [esp + nb112_fiyH2], xmm1 movlpd [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] addsd xmm0, [esp + nb112_fjxO] addsd xmm1, [esp + nb112_fjyO] addsd xmm2, [esp + nb112_fjzO] addsd xmm3, [esp + nb112_fjxH1] addsd xmm4, [esp + nb112_fjyH1]
⌨️ 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?