nb_kernel314_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,193 行 · 第 1/5 页
S
2,193 行
movsd [esp + nb314_jzH2], xmm4 movsd [esp + nb314_jyM], xmm5 unpckhpd xmm0, xmm0 unpckhpd xmm1, xmm1 unpckhpd xmm2, xmm2 unpckhpd xmm3, xmm3 unpckhpd xmm4, xmm4 unpckhpd xmm5, xmm5 movsd [esp + nb314_jyO], xmm0 movsd [esp + nb314_jxH1], xmm1 movsd [esp + nb314_jzH1], xmm2 movsd [esp + nb314_jyH2], xmm3 movsd [esp + nb314_jxM], xmm4 movsd [esp + nb314_jzM], xmm5 ;# start calculating pairwise distances movapd xmm0, [esp + nb314_ixO] movapd xmm1, [esp + nb314_iyO] movapd xmm2, [esp + nb314_izO] movapd xmm3, [esp + nb314_ixH1] movapd xmm4, [esp + nb314_iyH1] movapd xmm5, [esp + nb314_izH1] subsd xmm0, [esp + nb314_jxO] subsd xmm1, [esp + nb314_jyO] subsd xmm2, [esp + nb314_jzO] subsd xmm3, [esp + nb314_jxH1] subsd xmm4, [esp + nb314_jyH1] subsd xmm5, [esp + nb314_jzH1] movapd [esp + nb314_dxOO], xmm0 movapd [esp + nb314_dyOO], xmm1 movapd [esp + nb314_dzOO], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb314_dxH1H1], xmm3 movapd [esp + nb314_dyH1H1], xmm4 movapd [esp + nb314_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 + nb314_rsqOO], xmm0 movapd [esp + nb314_rsqH1H1], xmm3 movapd xmm0, [esp + nb314_ixH1] movapd xmm1, [esp + nb314_iyH1] movapd xmm2, [esp + nb314_izH1] movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 subsd xmm0, [esp + nb314_jxH2] subsd xmm1, [esp + nb314_jyH2] subsd xmm2, [esp + nb314_jzH2] subsd xmm3, [esp + nb314_jxM] subsd xmm4, [esp + nb314_jyM] subsd xmm5, [esp + nb314_jzM] movapd [esp + nb314_dxH1H2], xmm0 movapd [esp + nb314_dyH1H2], xmm1 movapd [esp + nb314_dzH1H2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb314_dxH1M], xmm3 movapd [esp + nb314_dyH1M], xmm4 movapd [esp + nb314_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 + nb314_rsqH1H2], xmm0 movapd [esp + nb314_rsqH1M], xmm3 movapd xmm0, [esp + nb314_ixH2] movapd xmm1, [esp + nb314_iyH2] movapd xmm2, [esp + nb314_izH2] movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 subsd xmm0, [esp + nb314_jxH1] subsd xmm1, [esp + nb314_jyH1] subsd xmm2, [esp + nb314_jzH1] subsd xmm3, [esp + nb314_jxH2] subsd xmm4, [esp + nb314_jyH2] subsd xmm5, [esp + nb314_jzH2] movapd [esp + nb314_dxH2H1], xmm0 movapd [esp + nb314_dyH2H1], xmm1 movapd [esp + nb314_dzH2H1], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb314_dxH2H2], xmm3 movapd [esp + nb314_dyH2H2], xmm4 movapd [esp + nb314_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 + nb314_rsqH2H1], xmm0 movapd [esp + nb314_rsqH2H2], xmm3 movapd xmm0, [esp + nb314_ixH2] movapd xmm1, [esp + nb314_iyH2] movapd xmm2, [esp + nb314_izH2] movapd xmm3, [esp + nb314_ixM] movapd xmm4, [esp + nb314_iyM] movapd xmm5, [esp + nb314_izM] subsd xmm0, [esp + nb314_jxM] subsd xmm1, [esp + nb314_jyM] subsd xmm2, [esp + nb314_jzM] subsd xmm3, [esp + nb314_jxH1] subsd xmm4, [esp + nb314_jyH1] subsd xmm5, [esp + nb314_jzH1] movapd [esp + nb314_dxH2M], xmm0 movapd [esp + nb314_dyH2M], xmm1 movapd [esp + nb314_dzH2M], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb314_dxMH1], xmm3 movapd [esp + nb314_dyMH1], xmm4 movapd [esp + nb314_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 + nb314_rsqH2M], xmm0 movapd [esp + nb314_rsqMH1], xmm4 movapd xmm0, [esp + nb314_ixM] movapd xmm1, [esp + nb314_iyM] movapd xmm2, [esp + nb314_izM] movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 subsd xmm0, [esp + nb314_jxH2] subsd xmm1, [esp + nb314_jyH2] subsd xmm2, [esp + nb314_jzH2] subsd xmm3, [esp + nb314_jxM] subsd xmm4, [esp + nb314_jyM] subsd xmm5, [esp + nb314_jzM] movapd [esp + nb314_dxMH2], xmm0 movapd [esp + nb314_dyMH2], xmm1 movapd [esp + nb314_dzMH2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb314_dxMM], xmm3 movapd [esp + nb314_dyMM], xmm4 movapd [esp + nb314_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 + nb314_rsqMH2], xmm0 movapd [esp + nb314_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 + nb314_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 + nb314_half] ;# iter1 mulsd xmm7, [esp + nb314_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 + nb314_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 + nb314_half] ;# rinv mulsd xmm5, [esp + nb314_half] ;# rinv movapd [esp + nb314_rinvMH2], xmm1 movapd [esp + nb314_rinvMM], xmm5 movapd xmm0, [esp + nb314_rsqOO] movapd xmm4, [esp + nb314_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 + nb314_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 + nb314_half] ;# iter1 of mulsd xmm7, [esp + nb314_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 + nb314_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 + nb314_half] ;# rinv mulsd xmm5, [esp + nb314_half] ;# rinv mulpd xmm1, xmm1 movapd [esp + nb314_rinvsqOO], xmm1 movapd [esp + nb314_rinvH1H1], xmm5 movapd xmm0, [esp + nb314_rsqH1H2] movapd xmm4, [esp + nb314_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 + nb314_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 + nb314_half] ;# iter1 mulsd xmm7, [esp + nb314_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 + nb314_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 + nb314_half] ;# rinv mulsd xmm5, [esp + nb314_half] ;# rinv movapd [esp + nb314_rinvH1H2], xmm1 movapd [esp + nb314_rinvH1M], xmm5 movapd xmm0, [esp + nb314_rsqH2H1] movapd xmm4, [esp + nb314_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 + nb314_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 + nb314_half] ;# iter1a mulsd xmm7, [esp + nb314_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 + nb314_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 + nb314_half] ;# rinv mulsd xmm5, [esp + nb314_half] ;# rinv movapd [esp + nb314_rinvH2H1], xmm1 movapd [esp + nb314_rinvH2H2], xmm5 movapd xmm0, [esp + nb314_rsqMH1] movapd xmm4, [esp + nb314_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 + nb314_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 + nb314_half] ;# iter1a mulsd xmm7, [esp + nb314_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 + nb314_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 + nb314_half] ;# rinv mulsd xmm5, [esp + nb314_half] ;# rinv movapd [esp + nb314_rinvMH1], xmm1 movapd [esp + nb314_rinvH2M], xmm5 ;# start with OO interaction movsd xmm0, [esp + nb314_rinvsqOO] ;# xmm0=rinvsq movapd xmm1, xmm0 mulsd xmm1, xmm1 ;# rinv4 mulsd xmm1, xmm0 ;#rinvsix movsd xmm2, xmm1 mulsd xmm2, xmm2 ;# rinvtwelve mulsd xmm1, [esp + nb314_c6] mulsd xmm2, [esp + nb314_c12] movsd xmm3, xmm2 subsd xmm3, xmm1 ;# Vvdw=Vvdw12-Vvdw6 addsd xmm3, [esp + nb314_Vvdwtot] mulsd xmm1, [esp + nb314_six] mulsd xmm2, [esp + nb314_twelve] subsd xmm2, xmm1 mulsd xmm2, xmm0 movsd [esp + nb314_Vvdwtot], xmm3 movapd xmm0, xmm2 movapd xmm1, xmm2 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb314_dxOO] mulsd xmm1, [esp + nb314_dyOO] mulsd xmm2, [esp + nb314_dzOO] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb314_fixO] addsd xmm1, [esp + nb314_fiyO] addsd xmm2, [esp + nb314_fizO] movsd [esp + nb314_fjxO], xmm3 movsd [esp + nb314_fjyO], xmm4 movsd [esp + nb314_fjzO], xmm5 movsd [esp + nb314_fixO], xmm0 movsd [esp + nb314_fiyO], xmm1 movsd [esp + nb314_fizO], xmm2 ;# H1-H1 interaction movapd xmm0, [esp + nb314_rinvH1H1] movapd xmm1, xmm0 mulsd xmm1, [esp + nb314_rsqH1H1] ;# xmm1=r mulsd xmm1, [esp + nb314_tsc] movd mm0, eax cvttsd2si eax, xmm1 ;# mm6 = lu idx cvtsi2sd xmm6, eax subsd xmm1, xmm6 ;# xmm1=eps movapd xmm2, xmm1 mulsd xmm2, xmm2 ;# xmm2=eps2 shl eax, 2 ;# idx *= 4 mov esi, [ebp + nb314_VFtab] movsd xmm4, [esi + eax*8] ;# Y1 movsd xmm5, [esi + eax*8 + 8] ;# F1 movsd xmm6, [esi + eax*8 + 16] ;# G1 movsd xmm7, [esi + eax*8 + 24] ;# H1
⌨️ 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?