nb_kernel104_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
mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 movapd [esp + nb104_rsqMM], xmm0 movapd [esp + nb104_rsqMH1], xmm3 movapd xmm0, [esp + nb104_ixM] movapd xmm1, [esp + nb104_iyM] movapd xmm2, [esp + nb104_izM] movapd xmm3, [esp + nb104_ixH1] movapd xmm4, [esp + nb104_iyH1] movapd xmm5, [esp + nb104_izH1] subpd xmm0, [esp + nb104_jxH2] subpd xmm1, [esp + nb104_jyH2] subpd xmm2, [esp + nb104_jzH2] subpd xmm3, [esp + nb104_jxM] subpd xmm4, [esp + nb104_jyM] subpd xmm5, [esp + nb104_jzM] movapd [esp + nb104_dxMH2], xmm0 movapd [esp + nb104_dyMH2], xmm1 movapd [esp + nb104_dzMH2], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb104_dxH1M], xmm3 movapd [esp + nb104_dyH1M], xmm4 movapd [esp + nb104_dzH1M], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 movapd [esp + nb104_rsqMH2], xmm0 movapd [esp + nb104_rsqH1M], xmm3 movapd xmm0, [esp + nb104_ixH1] movapd xmm1, [esp + nb104_iyH1] movapd xmm2, [esp + nb104_izH1] movapd xmm3, [esp + nb104_ixH1] movapd xmm4, [esp + nb104_iyH1] movapd xmm5, [esp + nb104_izH1] subpd xmm0, [esp + nb104_jxH1] subpd xmm1, [esp + nb104_jyH1] subpd xmm2, [esp + nb104_jzH1] subpd xmm3, [esp + nb104_jxH2] subpd xmm4, [esp + nb104_jyH2] subpd xmm5, [esp + nb104_jzH2] movapd [esp + nb104_dxH1H1], xmm0 movapd [esp + nb104_dyH1H1], xmm1 movapd [esp + nb104_dzH1H1], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb104_dxH1H2], xmm3 movapd [esp + nb104_dyH1H2], xmm4 movapd [esp + nb104_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 + nb104_rsqH1H1], xmm0 movapd [esp + nb104_rsqH1H2], xmm3 movapd xmm0, [esp + nb104_ixH2] movapd xmm1, [esp + nb104_iyH2] movapd xmm2, [esp + nb104_izH2] movapd xmm3, [esp + nb104_ixH2] movapd xmm4, [esp + nb104_iyH2] movapd xmm5, [esp + nb104_izH2] subpd xmm0, [esp + nb104_jxM] subpd xmm1, [esp + nb104_jyM] subpd xmm2, [esp + nb104_jzM] subpd xmm3, [esp + nb104_jxH1] subpd xmm4, [esp + nb104_jyH1] subpd xmm5, [esp + nb104_jzH1] movapd [esp + nb104_dxH2M], xmm0 movapd [esp + nb104_dyH2M], xmm1 movapd [esp + nb104_dzH2M], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb104_dxH2H1], xmm3 movapd [esp + nb104_dyH2H1], xmm4 movapd [esp + nb104_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 + nb104_rsqH2M], xmm0 movapd [esp + nb104_rsqH2H1], xmm4 movapd xmm0, [esp + nb104_ixH2] movapd xmm1, [esp + nb104_iyH2] movapd xmm2, [esp + nb104_izH2] subpd xmm0, [esp + nb104_jxH2] subpd xmm1, [esp + nb104_jyH2] subpd xmm2, [esp + nb104_jzH2] movapd [esp + nb104_dxH2H2], xmm0 movapd [esp + nb104_dyH2H2], xmm1 movapd [esp + nb104_dzH2H2], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 addpd xmm0, xmm1 addpd xmm0, xmm2 movapd [esp + nb104_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 + nb104_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 + nb104_half] ;# iter1 mulpd xmm7, [esp + nb104_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 + nb104_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 + nb104_half] ;# rinv mulpd xmm5, [esp + nb104_half] ;# rinv movapd [esp + nb104_rinvH2H2], xmm1 movapd [esp + nb104_rinvH2H1], xmm5 movapd xmm0, [esp + nb104_rsqMM] movapd xmm4, [esp + nb104_rsqMH1] 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 + nb104_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 + nb104_half] ;# iter1 of mulpd xmm7, [esp + nb104_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 + nb104_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 + nb104_half] ;# rinv mulpd xmm5, [esp + nb104_half] ;# rinv movapd [esp + nb104_rinvMM], xmm1 movapd [esp + nb104_rinvMH1], xmm5 movapd xmm0, [esp + nb104_rsqMH2] movapd xmm4, [esp + nb104_rsqH1M] 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 + nb104_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 + nb104_half] ;# iter1 mulpd xmm7, [esp + nb104_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 + nb104_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 + nb104_half] ;# rinv mulpd xmm5, [esp + nb104_half] ;# rinv movapd [esp + nb104_rinvMH2], xmm1 movapd [esp + nb104_rinvH1M], xmm5 movapd xmm0, [esp + nb104_rsqH1H1] movapd xmm4, [esp + nb104_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 + nb104_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 + nb104_half] ;# iter1a mulpd xmm7, [esp + nb104_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 + nb104_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 + nb104_half] ;# rinv mulpd xmm5, [esp + nb104_half] ;# rinv movapd [esp + nb104_rinvH1H1], xmm1 movapd [esp + nb104_rinvH1H2], xmm5 movapd xmm0, [esp + nb104_rsqH2M] cvtpd2ps xmm1, xmm0 rsqrtps xmm1, xmm1 cvtps2pd xmm1, xmm1 movapd xmm2, xmm1 ;# copy of luA mulpd xmm1, xmm1 ;# luA*luA movapd xmm3, [esp + nb104_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 + nb104_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA mulpd xmm3, xmm3 ;# luA*luA movapd xmm1, [esp + nb104_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 + nb104_half] ;# rinv movapd [esp + nb104_rinvH2M], xmm1 ;# start with MM interaction movapd xmm0, [esp + nb104_rinvMM] movapd xmm7, xmm0 mulpd xmm0, xmm0 ;# rinvsq mulpd xmm7, [esp + nb104_qqMM] mulpd xmm0, xmm7 addpd xmm7, [esp + nb104_vctot] movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb104_dxMM] mulpd xmm1, [esp + nb104_dyMM] mulpd xmm2, [esp + nb104_dzMM] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb104_fixM] addpd xmm1, [esp + nb104_fiyM] addpd xmm2, [esp + nb104_fizM] movapd [esp + nb104_fjxM], xmm3 movapd [esp + nb104_fjyM], xmm4 movapd [esp + nb104_fjzM], xmm5 movapd [esp + nb104_fixM], xmm0 movapd [esp + nb104_fiyM], xmm1 movapd [esp + nb104_fizM], xmm2 ;# M-H1 interaction movapd xmm0, [esp + nb104_rinvMH1] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb104_qqMH] mulpd xmm0, xmm1 ;# fsMH1 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 + nb104_dxMH1] mulpd xmm1, [esp + nb104_dyMH1] mulpd xmm2, [esp + nb104_dzMH1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb104_fixM] addpd xmm1, [esp + nb104_fiyM] addpd xmm2, [esp + nb104_fizM] movapd [esp + nb104_fjxH1], xmm3 movapd [esp + nb104_fjyH1], xmm4 movapd [esp + nb104_fjzH1], xmm5 movapd [esp + nb104_fixM], xmm0 movapd [esp + nb104_fiyM], xmm1 movapd [esp + nb104_fizM], xmm2 ;# M-H2 interaction movapd xmm0, [esp + nb104_rinvMH2] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb104_qqMH] mulpd xmm0, xmm1 ;# fsMH2 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 + nb104_dxMH2] mulpd xmm1, [esp + nb104_dyMH2] mulpd xmm2, [esp + nb104_dzMH2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb104_fixM] addpd xmm1, [esp + nb104_fiyM] addpd xmm2, [esp + nb104_fizM] movapd [esp + nb104_fjxH2], xmm3 movapd [esp + nb104_fjyH2], xmm4 movapd [esp + nb104_fjzH2], xmm5 movapd [esp + nb104_fixM], xmm0 movapd [esp + nb104_fiyM], xmm1 movapd [esp + nb104_fizM], xmm2 ;# H1-M interaction movapd xmm0, [esp + nb104_rinvH1M] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb104_qqMH] mulpd xmm0, xmm1 ;# fsH1M addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb104_fjxM] movapd xmm4, [esp + nb104_fjyM] movapd xmm5, [esp + nb104_fjzM] mulpd xmm0, [esp + nb104_dxH1M] mulpd xmm1, [esp + nb104_dyH1M] mulpd xmm2, [esp + nb104_dzH1M] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb104_fixH1] addpd xmm1, [esp + nb104_fiyH1] addpd xmm2, [esp + nb104_fizH1] movapd [esp + nb104_fjxM], xmm3 movapd [esp + nb104_fjyM], xmm4 movapd [esp + nb104_fjzM], xmm5 movapd [esp + nb104_fixH1], xmm0 movapd [esp + nb104_fiyH1], xmm1 movapd [esp + nb104_fizH1], xmm2 ;# H1-H1 interaction movapd xmm0, [esp + nb104_rinvH1H1] movapd xmm1, xmm0 mulpd xmm0, xmm0 mulpd xmm1, [esp + nb104_qqHH] mulpd xmm0, xmm1 ;# fsH1H1 addpd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb104_fjxH1] movapd xmm4, [esp + nb104_fjyH1] movapd xmm5, [esp + nb104_fjzH1] mulpd xmm0, [esp + nb104_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?