nb_kernel114_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,222 行 · 第 1/5 页
S
2,222 行
movsd [esp + nb114_jxM], xmm4 movsd [esp + nb114_jzM], xmm5 ;# start calculating pairwise distances movapd xmm0, [esp + nb114_ixO] movapd xmm1, [esp + nb114_iyO] movapd xmm2, [esp + nb114_izO] movapd xmm3, [esp + nb114_ixH1] movapd xmm4, [esp + nb114_iyH1] movapd xmm5, [esp + nb114_izH1] subsd xmm0, [esp + nb114_jxO] subsd xmm1, [esp + nb114_jyO] subsd xmm2, [esp + nb114_jzO] subsd xmm3, [esp + nb114_jxH1] subsd xmm4, [esp + nb114_jyH1] subsd xmm5, [esp + nb114_jzH1] movapd [esp + nb114_dxOO], xmm0 movapd [esp + nb114_dyOO], xmm1 movapd [esp + nb114_dzOO], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb114_dxH1H1], xmm3 movapd [esp + nb114_dyH1H1], xmm4 movapd [esp + nb114_dzH1H1], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd xmm3, xmm5 movapd [esp + nb114_rsqOO], xmm0 movapd [esp + nb114_rsqH1H1], xmm3 movapd xmm0, [esp + nb114_ixH1] movapd xmm1, [esp + nb114_iyH1] movapd xmm2, [esp + nb114_izH1] movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 subsd xmm0, [esp + nb114_jxH2] subsd xmm1, [esp + nb114_jyH2] subsd xmm2, [esp + nb114_jzH2] subsd xmm3, [esp + nb114_jxM] subsd xmm4, [esp + nb114_jyM] subsd xmm5, [esp + nb114_jzM] movapd [esp + nb114_dxH1H2], xmm0 movapd [esp + nb114_dyH1H2], xmm1 movapd [esp + nb114_dzH1H2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb114_dxH1M], xmm3 movapd [esp + nb114_dyH1M], xmm4 movapd [esp + nb114_dzH1M], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd xmm3, xmm5 movapd [esp + nb114_rsqH1H2], xmm0 movapd [esp + nb114_rsqH1M], xmm3 movapd xmm0, [esp + nb114_ixH2] movapd xmm1, [esp + nb114_iyH2] movapd xmm2, [esp + nb114_izH2] movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 subsd xmm0, [esp + nb114_jxH1] subsd xmm1, [esp + nb114_jyH1] subsd xmm2, [esp + nb114_jzH1] subsd xmm3, [esp + nb114_jxH2] subsd xmm4, [esp + nb114_jyH2] subsd xmm5, [esp + nb114_jzH2] movapd [esp + nb114_dxH2H1], xmm0 movapd [esp + nb114_dyH2H1], xmm1 movapd [esp + nb114_dzH2H1], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb114_dxH2H2], xmm3 movapd [esp + nb114_dyH2H2], xmm4 movapd [esp + nb114_dzH2H2], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd xmm3, xmm5 movapd [esp + nb114_rsqH2H1], xmm0 movapd [esp + nb114_rsqH2H2], xmm3 movapd xmm0, [esp + nb114_ixH2] movapd xmm1, [esp + nb114_iyH2] movapd xmm2, [esp + nb114_izH2] movapd xmm3, [esp + nb114_ixM] movapd xmm4, [esp + nb114_iyM] movapd xmm5, [esp + nb114_izM] subsd xmm0, [esp + nb114_jxM] subsd xmm1, [esp + nb114_jyM] subsd xmm2, [esp + nb114_jzM] subsd xmm3, [esp + nb114_jxH1] subsd xmm4, [esp + nb114_jyH1] subsd xmm5, [esp + nb114_jzH1] movapd [esp + nb114_dxH2M], xmm0 movapd [esp + nb114_dyH2M], xmm1 movapd [esp + nb114_dzH2M], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb114_dxMH1], xmm3 movapd [esp + nb114_dyMH1], xmm4 movapd [esp + nb114_dzMH1], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm4, xmm3 addsd xmm4, xmm5 movapd [esp + nb114_rsqH2M], xmm0 movapd [esp + nb114_rsqMH1], xmm4 movapd xmm0, [esp + nb114_ixM] movapd xmm1, [esp + nb114_iyM] movapd xmm2, [esp + nb114_izM] movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 subsd xmm0, [esp + nb114_jxH2] subsd xmm1, [esp + nb114_jyH2] subsd xmm2, [esp + nb114_jzH2] subsd xmm3, [esp + nb114_jxM] subsd xmm4, [esp + nb114_jyM] subsd xmm5, [esp + nb114_jzM] movapd [esp + nb114_dxMH2], xmm0 movapd [esp + nb114_dyMH2], xmm1 movapd [esp + nb114_dzMH2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb114_dxMM], xmm3 movapd [esp + nb114_dyMM], xmm4 movapd [esp + nb114_dzMM], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm4, xmm3 addsd xmm4, xmm5 movapd [esp + nb114_rsqMH2], xmm0 movapd [esp + nb114_rsqMM], xmm4 ;# Invsqrt form rsq M-H2 (rsq in xmm0) and MM (rsq in xmm4) cvtsd2ss xmm1, xmm0 cvtsd2ss xmm5, xmm4 rsqrtss xmm1, xmm1 rsqrtss xmm5, xmm5 cvtss2sd xmm1, xmm1 ;# luA cvtss2sd xmm5, xmm5 ;# luB 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 + nb114_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 + nb114_half] ;# iter1 mulsd xmm7, [esp + nb114_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 + nb114_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 + nb114_half] ;# rinv mulsd xmm5, [esp + nb114_half] ;# rinv movapd [esp + nb114_rinvMH2], xmm1 movapd [esp + nb114_rinvMM], xmm5 movapd xmm0, [esp + nb114_rsqOO] movapd xmm4, [esp + nb114_rsqH1H1] 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 + nb114_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 + nb114_half] ;# iter1 of mulsd xmm7, [esp + nb114_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 + nb114_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 + nb114_half] ;# rinv mulsd xmm5, [esp + nb114_half] ;# rinv mulpd xmm1, xmm1 movapd [esp + nb114_rinvsqOO], xmm1 movapd [esp + nb114_rinvH1H1], xmm5 movapd xmm0, [esp + nb114_rsqH1H2] movapd xmm4, [esp + nb114_rsqH1M] 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 + nb114_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 + nb114_half] ;# iter1 mulsd xmm7, [esp + nb114_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 + nb114_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 + nb114_half] ;# rinv mulsd xmm5, [esp + nb114_half] ;# rinv movapd [esp + nb114_rinvH1H2], xmm1 movapd [esp + nb114_rinvH1M], xmm5 movapd xmm0, [esp + nb114_rsqH2H1] movapd xmm4, [esp + nb114_rsqH2H2] 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 + nb114_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 + nb114_half] ;# iter1a mulsd xmm7, [esp + nb114_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 + nb114_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 + nb114_half] ;# rinv mulsd xmm5, [esp + nb114_half] ;# rinv movapd [esp + nb114_rinvH2H1], xmm1 movapd [esp + nb114_rinvH2H2], xmm5 movapd xmm0, [esp + nb114_rsqMH1] movapd xmm4, [esp + nb114_rsqH2M] 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 + nb114_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 + nb114_half] ;# iter1a mulsd xmm7, [esp + nb114_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 + nb114_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 + nb114_half] ;# rinv mulsd xmm5, [esp + nb114_half] ;# rinv movapd [esp + nb114_rinvMH1], xmm1 movapd [esp + nb114_rinvH2M], xmm5 ;# start with OO interaction movsd xmm0, [esp + nb114_rinvsqOO] ;# xmm0=rinvsq movapd xmm1, xmm0 mulsd xmm1, xmm1 ;# rinv4 mulsd xmm1, xmm0 ;#rinvsix movsd xmm2, xmm1 mulsd xmm2, xmm2 ;# rinvtwelve mulsd xmm1, [esp + nb114_c6] mulsd xmm2, [esp + nb114_c12] movsd xmm3, xmm2 subsd xmm3, xmm1 ;# Vvdw=Vvdw12-Vvdw6 addsd xmm3, [esp + nb114_Vvdwtot] mulsd xmm1, [esp + nb114_six] mulsd xmm2, [esp + nb114_twelve] subsd xmm2, xmm1 mulsd xmm2, xmm0 movsd [esp + nb114_Vvdwtot], xmm3 movapd xmm0, xmm2 movapd xmm1, xmm2 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb114_dxOO] mulsd xmm1, [esp + nb114_dyOO] mulsd xmm2, [esp + nb114_dzOO] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb114_fixO] addsd xmm1, [esp + nb114_fiyO] addsd xmm2, [esp + nb114_fizO] movsd [esp + nb114_fjxO], xmm3 movsd [esp + nb114_fjyO], xmm4 movsd [esp + nb114_fjzO], xmm5 movsd [esp + nb114_fixO], xmm0 movsd [esp + nb114_fiyO], xmm1 movsd [esp + nb114_fizO], xmm2 ;# H1-H1 interaction movsd xmm0, [esp + nb114_rinvH1H1] movsd xmm7, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm7, [esp + nb114_qqHH] mulsd xmm0, xmm7 addsd xmm7, [esp + nb114_vctot] movsd [esp + nb114_vctot], xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb114_dxH1H1] mulsd xmm1, [esp + nb114_dyH1H1] mulsd xmm2, [esp + nb114_dzH1H1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb114_fixH1] addsd xmm1, [esp + nb114_fiyH1] addsd xmm2, [esp + nb114_fizH1] movsd [esp + nb114_fjxH1], xmm3 movsd [esp + nb114_fjyH1], xmm4 movsd [esp + nb114_fjzH1], xmm5 movsd [esp + nb114_fixH1], xmm0 movsd [esp + nb114_fiyH1], xmm1 movsd [esp + nb114_fizH1], xmm2 ;# H1-H2 interaction movsd xmm0, [esp + nb114_rinvH1H2] movsd xmm7, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm7, [esp + nb114_qqHH] ;# vcoul mulsd xmm0, xmm7 addsd xmm7, [esp + nb114_vctot]
⌨️ 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?