nb_kernel232_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,152 行 · 第 1/5 页
S
2,152 行
mov esi, [ebp + nb232_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 + nb232_jxO], xmm2 movapd [esp + nb232_jyO], xmm3 movapd [esp + nb232_jzO], xmm4 movapd [esp + nb232_jxH1], xmm5 movapd [esp + nb232_jyH1], xmm6 movapd [esp + nb232_jzH1], xmm7 movlpd xmm2, [esi + eax*8 + 48] movlpd xmm3, [esi + eax*8 + 56] movlpd xmm4, [esi + eax*8 + 64] movapd [esp + nb232_jxH2], xmm2 movapd [esp + nb232_jyH2], xmm3 movapd [esp + nb232_jzH2], xmm4 movapd xmm0, [esp + nb232_ixO] movapd xmm1, [esp + nb232_iyO] movapd xmm2, [esp + nb232_izO] movapd xmm3, [esp + nb232_ixO] movapd xmm4, [esp + nb232_iyO] movapd xmm5, [esp + nb232_izO] subsd xmm0, [esp + nb232_jxO] subsd xmm1, [esp + nb232_jyO] subsd xmm2, [esp + nb232_jzO] subsd xmm3, [esp + nb232_jxH1] subsd xmm4, [esp + nb232_jyH1] subsd xmm5, [esp + nb232_jzH1] movapd [esp + nb232_dxOO], xmm0 movapd [esp + nb232_dyOO], xmm1 movapd [esp + nb232_dzOO], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb232_dxOH1], xmm3 movapd [esp + nb232_dyOH1], xmm4 movapd [esp + nb232_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 + nb232_rsqOO], xmm0 movapd [esp + nb232_rsqOH1], xmm3 movapd xmm0, [esp + nb232_ixO] movapd xmm1, [esp + nb232_iyO] movapd xmm2, [esp + nb232_izO] movapd xmm3, [esp + nb232_ixH1] movapd xmm4, [esp + nb232_iyH1] movapd xmm5, [esp + nb232_izH1] subsd xmm0, [esp + nb232_jxH2] subsd xmm1, [esp + nb232_jyH2] subsd xmm2, [esp + nb232_jzH2] subsd xmm3, [esp + nb232_jxO] subsd xmm4, [esp + nb232_jyO] subsd xmm5, [esp + nb232_jzO] movapd [esp + nb232_dxOH2], xmm0 movapd [esp + nb232_dyOH2], xmm1 movapd [esp + nb232_dzOH2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb232_dxH1O], xmm3 movapd [esp + nb232_dyH1O], xmm4 movapd [esp + nb232_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 + nb232_rsqOH2], xmm0 movapd [esp + nb232_rsqH1O], xmm3 movapd xmm0, [esp + nb232_ixH1] movapd xmm1, [esp + nb232_iyH1] movapd xmm2, [esp + nb232_izH1] movapd xmm3, [esp + nb232_ixH1] movapd xmm4, [esp + nb232_iyH1] movapd xmm5, [esp + nb232_izH1] subsd xmm0, [esp + nb232_jxH1] subsd xmm1, [esp + nb232_jyH1] subsd xmm2, [esp + nb232_jzH1] subsd xmm3, [esp + nb232_jxH2] subsd xmm4, [esp + nb232_jyH2] subsd xmm5, [esp + nb232_jzH2] movapd [esp + nb232_dxH1H1], xmm0 movapd [esp + nb232_dyH1H1], xmm1 movapd [esp + nb232_dzH1H1], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb232_dxH1H2], xmm3 movapd [esp + nb232_dyH1H2], xmm4 movapd [esp + nb232_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 + nb232_rsqH1H1], xmm0 movapd [esp + nb232_rsqH1H2], xmm3 movapd xmm0, [esp + nb232_ixH2] movapd xmm1, [esp + nb232_iyH2] movapd xmm2, [esp + nb232_izH2] movapd xmm3, [esp + nb232_ixH2] movapd xmm4, [esp + nb232_iyH2] movapd xmm5, [esp + nb232_izH2] subsd xmm0, [esp + nb232_jxO] subsd xmm1, [esp + nb232_jyO] subsd xmm2, [esp + nb232_jzO] subsd xmm3, [esp + nb232_jxH1] subsd xmm4, [esp + nb232_jyH1] subsd xmm5, [esp + nb232_jzH1] movapd [esp + nb232_dxH2O], xmm0 movapd [esp + nb232_dyH2O], xmm1 movapd [esp + nb232_dzH2O], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb232_dxH2H1], xmm3 movapd [esp + nb232_dyH2H1], xmm4 movapd [esp + nb232_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 + nb232_rsqH2O], xmm0 movapd [esp + nb232_rsqH2H1], xmm4 movapd xmm0, [esp + nb232_ixH2] movapd xmm1, [esp + nb232_iyH2] movapd xmm2, [esp + nb232_izH2] subsd xmm0, [esp + nb232_jxH2] subsd xmm1, [esp + nb232_jyH2] subsd xmm2, [esp + nb232_jzH2] movapd [esp + nb232_dxH2H2], xmm0 movapd [esp + nb232_dyH2H2], xmm1 movapd [esp + nb232_dzH2H2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 addsd xmm0, xmm1 addsd xmm0, xmm2 movapd [esp + nb232_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 + nb232_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 + nb232_half] ;# iter1 mulsd xmm7, [esp + nb232_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 + nb232_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 + nb232_half] ;# rinv mulsd xmm5, [esp + nb232_half] ;# rinv movapd [esp + nb232_rinvH2H2], xmm1 movapd [esp + nb232_rinvH2H1], xmm5 movapd xmm0, [esp + nb232_rsqOO] movapd xmm4, [esp + nb232_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 + nb232_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 + nb232_half] ;# iter1 of mulsd xmm7, [esp + nb232_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 + nb232_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 + nb232_half] ;# rinv mulsd xmm5, [esp + nb232_half] ;# rinv movapd [esp + nb232_rinvOO], xmm1 movapd [esp + nb232_rinvOH1], xmm5 movapd xmm0, [esp + nb232_rsqOH2] movapd xmm4, [esp + nb232_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 + nb232_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 + nb232_half] ;# iter1 mulsd xmm7, [esp + nb232_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 + nb232_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 + nb232_half] ;# rinv mulsd xmm5, [esp + nb232_half] ;# rinv movapd [esp + nb232_rinvOH2], xmm1 movapd [esp + nb232_rinvH1O], xmm5 movapd xmm0, [esp + nb232_rsqH1H1] movapd xmm4, [esp + nb232_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 + nb232_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 + nb232_half] ;# iter1a mulsd xmm7, [esp + nb232_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 + nb232_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 + nb232_half] ;# rinv mulsd xmm5, [esp + nb232_half] ;# rinv movapd [esp + nb232_rinvH1H1], xmm1 movapd [esp + nb232_rinvH1H2], xmm5 movapd xmm0, [esp + nb232_rsqH2O] cvtsd2ss xmm1, xmm0 rsqrtss xmm1, xmm1 cvtss2sd xmm1, xmm1 movapd xmm2, xmm1 ;# copy of luA mulsd xmm1, xmm1 ;# luA*luA movapd xmm3, [esp + nb232_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 + nb232_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA mulsd xmm3, xmm3 ;# luA*luA movapd xmm1, [esp + nb232_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 + nb232_half] ;# rinv movapd [esp + nb232_rinvH2O], xmm1 ;# start with OO interaction movsd xmm0, [esp + nb232_rinvOO] movsd xmm7, xmm0 ;# xmm7=rinv movsd xmm5, [esp + nb232_krf] mulsd xmm5, [esp + nb232_rsqOO] ;# xmm5=krsq movsd xmm6, xmm5 ;# krsq addsd xmm6, xmm7 ;# xmm6=rinv+ krsq subsd xmm6, [esp + nb232_crf] ;# rinv+krsq-crf mulsd xmm6, [esp + nb232_qqOO] ;# xmm6=voul=qq*(rinv+ krsq-crf) mulsd xmm5, [esp + nb232_two] ;# 2*krsq subsd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulsd xmm7, [esp + nb232_qqOO] ;# xmm7 = qq*(rinv-2*krsq) mulsd xmm7, xmm0 addsd xmm6, [esp + nb232_vctot] movsd [esp + nb232_vctot], xmm6 movsd [esp + nb232_fstmp], xmm7 mulsd xmm7, xmm0 ;# LJ table interaction movsd xmm4, [esp + nb232_rsqOO] mulsd xmm4, xmm0 ;# xmm4=r mulsd xmm4, [esp + nb232_tsc] cvttsd2si ebx, xmm4 ;# mm6 = lu idx cvtsi2sd xmm5, ebx subsd xmm4, xmm5 movsd xmm1, xmm4 ;# xmm1=eps movsd xmm2, xmm1 mulsd xmm2, xmm2 ;# xmm2=eps2 shl ebx, 3 mov esi, [ebp + nb232_VFtab] ;# dispersion movlpd xmm4, [esi + ebx*8] ;# Y1 movhpd xmm4, [esi + ebx*8 + 8] ;# Y1 F1 movapd xmm5, xmm4 unpcklpd xmm4, xmm3 ;# Y1 Y2 unpckhpd xmm5, xmm3 ;# F1 F2 movlpd xmm6, [esi + ebx*8 + 16] ;# G1 movhpd xmm6, [esi + ebx*8 + 24] ;# G1 H1 movapd xmm7, xmm6 unpcklpd xmm6, xmm3 ;# G1 G2 unpckhpd xmm7, xmm3 ;# H1 H2 ;# dispersion table ready, in xmm4-xmm7 mulsd xmm6, xmm1 ;# xmm6=Geps mulsd xmm7, xmm2 ;# xmm7=Heps2 addsd xmm5, xmm6 addsd xmm5, xmm7 ;# xmm5=Fp mulsd xmm7, [esp + nb232_two] ;# two*Heps2 addsd xmm7, xmm6 addsd xmm7, xmm5 ;# xmm7=FF mulsd xmm5, xmm1 ;# xmm5=eps*Fp addsd xmm5, xmm4 ;# xmm5=VV movsd xmm4, [esp + nb232_c6] mulsd xmm7, xmm4 ;# fijD mulsd xmm5, xmm4 ;# Vvdw6 ;# put scalar force on stack Update Vvdwtot directly addsd xmm5, [esp + nb232_Vvdwtot] movsd xmm3, [esp + nb232_fstmp] mulsd xmm7, [esp + nb232_tsc] subsd xmm3, xmm7 movsd [esp + nb232_fstmp], xmm3 movsd [esp + nb232_Vvdwtot], xmm5 ;# repulsion movlpd xmm4, [esi + ebx*8 + 32] ;# Y1 movhpd xmm4, [esi + ebx*8 + 40] ;# Y1 F1 movapd xmm5, xmm4 unpcklpd xmm4, xmm3 ;# Y1 Y2 unpckhpd xmm5, xmm3 ;# F1 F2 movlpd xmm6, [esi + ebx*8 + 48] ;# G1 movhpd xmm6, [esi + ebx*8 + 56] ;# G1 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?