nb_kernel102_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,155 行 · 第 1/5 页
S
2,155 行
mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 movapd [esp + nb102_rsqOO], xmm0 movapd [esp + nb102_rsqOH1], xmm3 movapd xmm0, [esp + nb102_ixO] movapd xmm1, [esp + nb102_iyO] movapd xmm2, [esp + nb102_izO] movapd xmm3, [esp + nb102_ixH1] movapd xmm4, [esp + nb102_iyH1] movapd xmm5, [esp + nb102_izH1] subpd xmm0, [esp + nb102_jxH2] subpd xmm1, [esp + nb102_jyH2] subpd xmm2, [esp + nb102_jzH2] subpd xmm3, [esp + nb102_jxO] subpd xmm4, [esp + nb102_jyO] subpd xmm5, [esp + nb102_jzO] movapd [esp + nb102_dxOH2], xmm0 movapd [esp + nb102_dyOH2], xmm1 movapd [esp + nb102_dzOH2], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb102_dxH1O], xmm3 movapd [esp + nb102_dyH1O], xmm4 movapd [esp + nb102_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 + nb102_rsqOH2], xmm0 movapd [esp + nb102_rsqH1O], xmm3 movapd xmm0, [esp + nb102_ixH1] movapd xmm1, [esp + nb102_iyH1] movapd xmm2, [esp + nb102_izH1] movapd xmm3, [esp + nb102_ixH1] movapd xmm4, [esp + nb102_iyH1] movapd xmm5, [esp + nb102_izH1] subpd xmm0, [esp + nb102_jxH1] subpd xmm1, [esp + nb102_jyH1] subpd xmm2, [esp + nb102_jzH1] subpd xmm3, [esp + nb102_jxH2] subpd xmm4, [esp + nb102_jyH2] subpd xmm5, [esp + nb102_jzH2] movapd [esp + nb102_dxH1H1], xmm0 movapd [esp + nb102_dyH1H1], xmm1 movapd [esp + nb102_dzH1H1], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb102_dxH1H2], xmm3 movapd [esp + nb102_dyH1H2], xmm4 movapd [esp + nb102_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 + nb102_rsqH1H1], xmm0 movapd [esp + nb102_rsqH1H2], xmm3 movapd xmm0, [esp + nb102_ixH2] movapd xmm1, [esp + nb102_iyH2] movapd xmm2, [esp + nb102_izH2] movapd xmm3, [esp + nb102_ixH2] movapd xmm4, [esp + nb102_iyH2] movapd xmm5, [esp + nb102_izH2] subpd xmm0, [esp + nb102_jxO] subpd xmm1, [esp + nb102_jyO] subpd xmm2, [esp + nb102_jzO] subpd xmm3, [esp + nb102_jxH1] subpd xmm4, [esp + nb102_jyH1] subpd xmm5, [esp + nb102_jzH1] movapd [esp + nb102_dxH2O], xmm0 movapd [esp + nb102_dyH2O], xmm1 movapd [esp + nb102_dzH2O], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb102_dxH2H1], xmm3 movapd [esp + nb102_dyH2H1], xmm4 movapd [esp + nb102_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 + nb102_rsqH2O], xmm0 movapd [esp + nb102_rsqH2H1], xmm4 movapd xmm0, [esp + nb102_ixH2] movapd xmm1, [esp + nb102_iyH2] movapd xmm2, [esp + nb102_izH2] subpd xmm0, [esp + nb102_jxH2] subpd xmm1, [esp + nb102_jyH2] subpd xmm2, [esp + nb102_jzH2] movapd [esp + nb102_dxH2H2], xmm0 movapd [esp + nb102_dyH2H2], xmm1 movapd [esp + nb102_dzH2H2], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 addpd xmm0, xmm1 addpd xmm0, xmm2 movapd [esp + nb102_rsqH2H2], xmm0 ;# start doing invsqrt use rsq values in xmm0 (h2h2) , xmm4 (h2h1) 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 + nb102_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 + nb102_half] ;# iter1 mulpd xmm7, [esp + nb102_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 + nb102_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 + nb102_half] ;# rinv mulpd xmm5, [esp + nb102_half] ;# rinv movapd [esp + nb102_rinvH2H2], xmm1 movapd [esp + nb102_rinvH2H1], xmm5 movapd xmm0, [esp + nb102_rsqOO] movapd xmm4, [esp + nb102_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 + nb102_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 + nb102_half] ;# iter1 of mulpd xmm7, [esp + nb102_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 + nb102_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 + nb102_half] ;# rinv mulpd xmm5, [esp + nb102_half] ;# rinv movapd [esp + nb102_rinvOO], xmm1 movapd [esp + nb102_rinvOH1], xmm5 movapd xmm0, [esp + nb102_rsqOH2] movapd xmm4, [esp + nb102_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 + nb102_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 + nb102_half] ;# iter1 mulpd xmm7, [esp + nb102_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 + nb102_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 + nb102_half] ;# rinv mulpd xmm5, [esp + nb102_half] ;# rinv movapd [esp + nb102_rinvOH2], xmm1 movapd [esp + nb102_rinvH1O], xmm5 movapd xmm0, [esp + nb102_rsqH1H1] movapd xmm4, [esp + nb102_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 + nb102_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 + nb102_half] ;# iter1a mulpd xmm7, [esp + nb102_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 + nb102_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 + nb102_half] ;# rinv mulpd xmm5, [esp + nb102_half] ;# rinv movapd [esp + nb102_rinvH1H1], xmm1 movapd [esp + nb102_rinvH1H2], xmm5 movapd xmm0, [esp + nb102_rsqH2O] cvtpd2ps xmm1, xmm0 rsqrtps xmm1, xmm1 cvtps2pd xmm1, xmm1 movapd xmm2, xmm1 ;# copy of luA mulpd xmm1, xmm1 ;# luA*luA movapd xmm3, [esp + nb102_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 + nb102_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA mulpd xmm3, xmm3 ;# luA*luA movapd xmm1, [esp + nb102_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 + nb102_half] ;# rinv movapd [esp + nb102_rinvH2O], xmm1 ;# start with OO interaction movapd xmm0, [esp + nb102_rinvOO] movapd xmm7, xmm0 mulpd xmm0, xmm0 ;# rinvsq mulpd xmm7, [esp + nb102_qqOO] mulpd xmm0, xmm7 addpd xmm7, [esp + nb102_vctot] movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb102_dxOO] mulpd xmm1, [esp + nb102_dyOO] mulpd xmm2, [esp + nb102_dzOO] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb102_fixO] addpd xmm1, [esp + nb102_fiyO] addpd xmm2, [esp + nb102_fizO] movapd [esp + nb102_fjxO], xmm3 movapd [esp + nb102_fjyO], xmm4 movapd [esp + nb102_fjzO], xmm5 movapd [esp + nb102_fixO], xmm0 movapd [esp + nb102_fiyO], xmm1 movapd [esp + nb102_fizO], xmm2 ;# O-H1 interaction movapd xmm0, [esp + nb102_rinvOH1] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb102_qqOH] mulpd xmm0, xmm1 ;# fsOH1 addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb102_dxOH1] mulpd xmm1, [esp + nb102_dyOH1] mulpd xmm2, [esp + nb102_dzOH1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb102_fixO] addpd xmm1, [esp + nb102_fiyO] addpd xmm2, [esp + nb102_fizO] movapd [esp + nb102_fjxH1], xmm3 movapd [esp + nb102_fjyH1], xmm4 movapd [esp + nb102_fjzH1], xmm5 movapd [esp + nb102_fixO], xmm0 movapd [esp + nb102_fiyO], xmm1 movapd [esp + nb102_fizO], xmm2 ;# O-H2 interaction movapd xmm0, [esp + nb102_rinvOH2] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb102_qqOH] mulpd xmm0, xmm1 ;# fsOH2 addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb102_dxOH2] mulpd xmm1, [esp + nb102_dyOH2] mulpd xmm2, [esp + nb102_dzOH2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb102_fixO] addpd xmm1, [esp + nb102_fiyO] addpd xmm2, [esp + nb102_fizO] movapd [esp + nb102_fjxH2], xmm3 movapd [esp + nb102_fjyH2], xmm4 movapd [esp + nb102_fjzH2], xmm5 movapd [esp + nb102_fixO], xmm0 movapd [esp + nb102_fiyO], xmm1 movapd [esp + nb102_fizO], xmm2 ;# H1-O interaction movapd xmm0, [esp + nb102_rinvH1O] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb102_qqOH] mulpd xmm0, xmm1 ;# fsH1O addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb102_fjxO] movapd xmm4, [esp + nb102_fjyO] movapd xmm5, [esp + nb102_fjzO] mulpd xmm0, [esp + nb102_dxH1O] mulpd xmm1, [esp + nb102_dyH1O] mulpd xmm2, [esp + nb102_dzH1O] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb102_fixH1] addpd xmm1, [esp + nb102_fiyH1] addpd xmm2, [esp + nb102_fizH1] movapd [esp + nb102_fjxO], xmm3 movapd [esp + nb102_fjyO], xmm4 movapd [esp + nb102_fjzO], xmm5 movapd [esp + nb102_fixH1], xmm0 movapd [esp + nb102_fiyH1], xmm1 movapd [esp + nb102_fizH1], xmm2 ;# H1-H1 interaction movapd xmm0, [esp + nb102_rinvH1H1] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb102_qqHH] mulpd xmm0, xmm1 ;# fsH1H1 addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb102_fjxH1] movapd xmm4, [esp + nb102_fjyH1] movapd xmm5, [esp + nb102_fjzH1] mulpd xmm0, [esp + nb102_dxH1H1]
⌨️ 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?