nb_kernel212_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,141 行 · 第 1/5 页
S
2,141 行
movhpd xmm3, [esi + ebx*8 + 8] movhpd xmm4, [esi + ebx*8 + 16] movhpd xmm5, [esi + ebx*8 + 24] movhpd xmm6, [esi + ebx*8 + 32] movhpd xmm7, [esi + ebx*8 + 40] movapd [esp + nb212_jxO], xmm2 movapd [esp + nb212_jyO], xmm3 movapd [esp + nb212_jzO], xmm4 movapd [esp + nb212_jxH1], xmm5 movapd [esp + nb212_jyH1], xmm6 movapd [esp + nb212_jzH1], xmm7 movlpd xmm2, [esi + eax*8 + 48] movlpd xmm3, [esi + eax*8 + 56] movlpd xmm4, [esi + eax*8 + 64] movhpd xmm2, [esi + ebx*8 + 48] movhpd xmm3, [esi + ebx*8 + 56] movhpd xmm4, [esi + ebx*8 + 64] movapd [esp + nb212_jxH2], xmm2 movapd [esp + nb212_jyH2], xmm3 movapd [esp + nb212_jzH2], xmm4 movapd xmm0, [esp + nb212_ixO] movapd xmm1, [esp + nb212_iyO] movapd xmm2, [esp + nb212_izO] movapd xmm3, [esp + nb212_ixO] movapd xmm4, [esp + nb212_iyO] movapd xmm5, [esp + nb212_izO] subpd xmm0, [esp + nb212_jxO] subpd xmm1, [esp + nb212_jyO] subpd xmm2, [esp + nb212_jzO] subpd xmm3, [esp + nb212_jxH1] subpd xmm4, [esp + nb212_jyH1] subpd xmm5, [esp + nb212_jzH1] movapd [esp + nb212_dxOO], xmm0 movapd [esp + nb212_dyOO], xmm1 movapd [esp + nb212_dzOO], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb212_dxOH1], xmm3 movapd [esp + nb212_dyOH1], xmm4 movapd [esp + nb212_dzOH1], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 movapd [esp + nb212_rsqOO], xmm0 movapd [esp + nb212_rsqOH1], xmm3 movapd xmm0, [esp + nb212_ixO] movapd xmm1, [esp + nb212_iyO] movapd xmm2, [esp + nb212_izO] movapd xmm3, [esp + nb212_ixH1] movapd xmm4, [esp + nb212_iyH1] movapd xmm5, [esp + nb212_izH1] subpd xmm0, [esp + nb212_jxH2] subpd xmm1, [esp + nb212_jyH2] subpd xmm2, [esp + nb212_jzH2] subpd xmm3, [esp + nb212_jxO] subpd xmm4, [esp + nb212_jyO] subpd xmm5, [esp + nb212_jzO] movapd [esp + nb212_dxOH2], xmm0 movapd [esp + nb212_dyOH2], xmm1 movapd [esp + nb212_dzOH2], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb212_dxH1O], xmm3 movapd [esp + nb212_dyH1O], xmm4 movapd [esp + nb212_dzH1O], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 movapd [esp + nb212_rsqOH2], xmm0 movapd [esp + nb212_rsqH1O], xmm3 movapd xmm0, [esp + nb212_ixH1] movapd xmm1, [esp + nb212_iyH1] movapd xmm2, [esp + nb212_izH1] movapd xmm3, [esp + nb212_ixH1] movapd xmm4, [esp + nb212_iyH1] movapd xmm5, [esp + nb212_izH1] subpd xmm0, [esp + nb212_jxH1] subpd xmm1, [esp + nb212_jyH1] subpd xmm2, [esp + nb212_jzH1] subpd xmm3, [esp + nb212_jxH2] subpd xmm4, [esp + nb212_jyH2] subpd xmm5, [esp + nb212_jzH2] movapd [esp + nb212_dxH1H1], xmm0 movapd [esp + nb212_dyH1H1], xmm1 movapd [esp + nb212_dzH1H1], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb212_dxH1H2], xmm3 movapd [esp + nb212_dyH1H2], xmm4 movapd [esp + nb212_dzH1H2], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 movapd [esp + nb212_rsqH1H1], xmm0 movapd [esp + nb212_rsqH1H2], xmm3 movapd xmm0, [esp + nb212_ixH2] movapd xmm1, [esp + nb212_iyH2] movapd xmm2, [esp + nb212_izH2] movapd xmm3, [esp + nb212_ixH2] movapd xmm4, [esp + nb212_iyH2] movapd xmm5, [esp + nb212_izH2] subpd xmm0, [esp + nb212_jxO] subpd xmm1, [esp + nb212_jyO] subpd xmm2, [esp + nb212_jzO] subpd xmm3, [esp + nb212_jxH1] subpd xmm4, [esp + nb212_jyH1] subpd xmm5, [esp + nb212_jzH1] movapd [esp + nb212_dxH2O], xmm0 movapd [esp + nb212_dyH2O], xmm1 movapd [esp + nb212_dzH2O], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb212_dxH2H1], xmm3 movapd [esp + nb212_dyH2H1], xmm4 movapd [esp + nb212_dzH2H1], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm4, xmm3 addpd xmm4, xmm5 movapd [esp + nb212_rsqH2O], xmm0 movapd [esp + nb212_rsqH2H1], xmm4 movapd xmm0, [esp + nb212_ixH2] movapd xmm1, [esp + nb212_iyH2] movapd xmm2, [esp + nb212_izH2] subpd xmm0, [esp + nb212_jxH2] subpd xmm1, [esp + nb212_jyH2] subpd xmm2, [esp + nb212_jzH2] movapd [esp + nb212_dxH2H2], xmm0 movapd [esp + nb212_dyH2H2], xmm1 movapd [esp + nb212_dzH2H2], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 addpd xmm0, xmm1 addpd xmm0, xmm2 movapd [esp + nb212_rsqH2H2], xmm0 ;# start doing invsqrt use rsq values in xmm0, xmm4 cvtpd2ps xmm1, xmm0 cvtpd2ps xmm5, xmm4 rsqrtps xmm1, xmm1 rsqrtps xmm5, xmm5 cvtps2pd xmm1, xmm1 cvtps2pd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulpd xmm1, xmm1 ;# luA*luA mulpd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb212_three] mulpd xmm1, xmm0 ;# rsqA*luA*luA mulpd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subpd xmm3, xmm1 ;# 3-rsqA*luA*luA subpd xmm7, xmm5 ;# 3-rsqB*luB*luB mulpd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm3, [esp + nb212_half] ;# iter1 mulpd xmm7, [esp + nb212_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulpd xmm3, xmm3 ;# luA*luA mulpd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb212_three] mulpd xmm3, xmm0 ;# rsqA*luA*luA mulpd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subpd xmm1, xmm3 ;# 3-rsqA*luA*luA subpd xmm5, xmm7 ;# 3-rsqB*luB*luB mulpd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm1, [esp + nb212_half] ;# rinv mulpd xmm5, [esp + nb212_half] ;# rinv movapd [esp + nb212_rinvH2H2], xmm1 movapd [esp + nb212_rinvH2H1], xmm5 movapd xmm0, [esp + nb212_rsqOO] movapd xmm4, [esp + nb212_rsqOH1] cvtpd2ps xmm1, xmm0 cvtpd2ps xmm5, xmm4 rsqrtps xmm1, xmm1 rsqrtps xmm5, xmm5 cvtps2pd xmm1, xmm1 cvtps2pd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulpd xmm1, xmm1 ;# luA*luA mulpd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb212_three] mulpd xmm1, xmm0 ;# rsqA*luA*luA mulpd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subpd xmm3, xmm1 ;# 3-rsqA*luA*luA subpd xmm7, xmm5 ;# 3-rsqB*luB*luB mulpd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm3, [esp + nb212_half] ;# iter1 of mulpd xmm7, [esp + nb212_half] ;# iter1 of movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulpd xmm3, xmm3 ;# luA*luA mulpd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb212_three] mulpd xmm3, xmm0 ;# rsqA*luA*luA mulpd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subpd xmm1, xmm3 ;# 3-rsqA*luA*luA subpd xmm5, xmm7 ;# 3-rsqB*luB*luB mulpd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm1, [esp + nb212_half] ;# rinv mulpd xmm5, [esp + nb212_half] ;# rinv movapd [esp + nb212_rinvOO], xmm1 movapd [esp + nb212_rinvOH1], xmm5 movapd xmm0, [esp + nb212_rsqOH2] movapd xmm4, [esp + nb212_rsqH1O] cvtpd2ps xmm1, xmm0 cvtpd2ps xmm5, xmm4 rsqrtps xmm1, xmm1 rsqrtps xmm5, xmm5 cvtps2pd xmm1, xmm1 cvtps2pd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulpd xmm1, xmm1 ;# luA*luA mulpd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb212_three] mulpd xmm1, xmm0 ;# rsqA*luA*luA mulpd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subpd xmm3, xmm1 ;# 3-rsqA*luA*luA subpd xmm7, xmm5 ;# 3-rsqB*luB*luB mulpd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm3, [esp + nb212_half] ;# iter1 mulpd xmm7, [esp + nb212_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulpd xmm3, xmm3 ;# luA*luA mulpd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb212_three] mulpd xmm3, xmm0 ;# rsqA*luA*luA mulpd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subpd xmm1, xmm3 ;# 3-rsqA*luA*luA subpd xmm5, xmm7 ;# 3-rsqB*luB*luB mulpd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm1, [esp + nb212_half] ;# rinv mulpd xmm5, [esp + nb212_half] ;# rinv movapd [esp + nb212_rinvOH2], xmm1 movapd [esp + nb212_rinvH1O], xmm5 movapd xmm0, [esp + nb212_rsqH1H1] movapd xmm4, [esp + nb212_rsqH1H2] cvtpd2ps xmm1, xmm0 cvtpd2ps xmm5, xmm4 rsqrtps xmm1, xmm1 rsqrtps xmm5, xmm5 cvtps2pd xmm1, xmm1 cvtps2pd xmm5, xmm5 movapd xmm2, xmm1 ;# copy of luA movapd xmm6, xmm5 ;# copy of luB mulpd xmm1, xmm1 ;# luA*luA mulpd xmm5, xmm5 ;# luB*luB movapd xmm3, [esp + nb212_three] mulpd xmm1, xmm0 ;# rsqA*luA*luA mulpd xmm5, xmm4 ;# rsqB*luB*luB movapd xmm7, xmm3 subpd xmm3, xmm1 ;# 3-rsqA*luA*luA subpd xmm7, xmm5 ;# 3-rsqB*luB*luB mulpd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm7, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm3, [esp + nb212_half] ;# iter1a mulpd xmm7, [esp + nb212_half] ;# iter1b movapd xmm2, xmm3 ;# copy of luA movapd xmm6, xmm7 ;# copy of luB mulpd xmm3, xmm3 ;# luA*luA mulpd xmm7, xmm7 ;# luB*luB movapd xmm1, [esp + nb212_three] mulpd xmm3, xmm0 ;# rsqA*luA*luA mulpd xmm7, xmm4 ;# rsqB*luB*luB movapd xmm5, xmm1 subpd xmm1, xmm3 ;# 3-rsqA*luA*luA subpd xmm5, xmm7 ;# 3-rsqB*luB*luB mulpd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm5, xmm6 ;# luB*(3-rsqB*luB*luB) mulpd xmm1, [esp + nb212_half] ;# rinv mulpd xmm5, [esp + nb212_half] ;# rinv movapd [esp + nb212_rinvH1H1], xmm1 movapd [esp + nb212_rinvH1H2], xmm5 movapd xmm0, [esp + nb212_rsqH2O] cvtpd2ps xmm1, xmm0 rsqrtps xmm1, xmm1 cvtps2pd xmm1, xmm1 movapd xmm2, xmm1 ;# copy of luA mulpd xmm1, xmm1 ;# luA*luA movapd xmm3, [esp + nb212_three] mulpd xmm1, xmm0 ;# rsqA*luA*luA subpd xmm3, xmm1 ;# 3-rsqA*luA*luA mulpd xmm3, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm3, [esp + nb212_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA mulpd xmm3, xmm3 ;# luA*luA movapd xmm1, [esp + nb212_three] mulpd xmm3, xmm0 ;# rsqA*luA*luA subpd xmm1, xmm3 ;# 3-rsqA*luA*luA mulpd xmm1, xmm2 ;# luA*(3-rsqA*luA*luA) mulpd xmm1, [esp + nb212_half] ;# rinv movapd [esp + nb212_rinvH2O], xmm1 ;# start with OO interaction movapd xmm0, [esp + nb212_rinvOO] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb212_krf] mulpd xmm0, xmm0 ;# xmm0=rinvsq movapd xmm1, xmm0 mulpd xmm1, xmm0 ;# rinvsq*rinvsq mulpd xmm1, xmm0 ;# xmm1=rinvsix mulpd xmm5, [esp + nb212_rsqOO] ;# xmm5=krsq movapd xmm6, xmm5 ;# krsq addpd xmm6, xmm7 ;# xmm6=rinv+ krsq subpd xmm6, [esp + nb212_crf] ;# rinv+krsq-crf mulpd xmm6, [esp + nb212_qqOO] ;# xmm6=voul=qq*(rinv+ krsq-crf) mulpd xmm5, [esp + nb212_two] ;# 2*krsq subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb212_qqOO] ;# xmm7 = qq*(rinv-2*krsq) movapd xmm2, xmm1 ;# rinv6 mulpd xmm2, xmm2 ;# xmm2=rinvtwelve mulpd xmm1, [esp + nb212_c6] ;# c6*rinv6 mulpd xmm2, [esp + nb212_c12] ;# c12*rinv12 movapd xmm3, xmm2 ;# c12*rinv12 subpd xmm3, xmm1 ;# Vvdw12-Vvdw6 addpd xmm3, [esp + nb212_Vvdwtot] mulpd xmm1, [esp + nb212_six] ;# 6.0*Vvdw6 mulpd xmm2, [esp + nb212_twelve] ;# 12*Vvdw12 movapd [esp + nb212_Vvdwtot], xmm3 subpd xmm2, xmm1 ;# 12*Vvdw12-6*Vvdw6 addpd xmm2, xmm7 ;# 12*Vvdw12-6*Vvdw6+qq*(rinv-2*krsq) addpd xmm6, [esp + nb212_vctot] ;# local vctot summation variable mulpd xmm0, xmm2 ;# (12*Vvdw12-6*Vvdw6+qq*(rinv-2*krsq))*rinvsq movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb212_dxOO] mulpd xmm1, [esp + nb212_dyOO] mulpd xmm2, [esp + nb212_dzOO] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb212_fixO] addpd xmm1, [esp + nb212_fiyO] addpd xmm2, [esp + nb212_fizO] movapd [esp + nb212_fjxO], xmm3 movapd [esp + nb212_fjyO], xmm4 movapd [esp + nb212_fjzO], xmm5 movapd [esp + nb212_fixO], xmm0 movapd [esp + nb212_fiyO], xmm1 movapd [esp + nb212_fizO], xmm2 ;# O-H1 interaction movapd xmm0, [esp + nb212_rinvOH1] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb212_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb212_rsqOH1] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=rinv+ krsq mulpd xmm0, xmm0 subpd xmm4, [esp + nb212_crf] mulpd xmm4, [esp + nb212_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb212_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb212_qqOH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH1 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb212_dxOH1] mulpd xmm1, [esp + nb212_dyOH1] mulpd xmm2, [esp + nb212_dzOH1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2
⌨️ 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?