nb_kernel204_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,120 行 · 第 1/5 页
S
2,120 行
movapd xmm2, [esp + nb204_izM] movapd xmm3, [esp + nb204_ixM] movapd xmm4, [esp + nb204_iyM] movapd xmm5, [esp + nb204_izM] subpd xmm0, [esp + nb204_jxM] subpd xmm1, [esp + nb204_jyM] subpd xmm2, [esp + nb204_jzM] subpd xmm3, [esp + nb204_jxH1] subpd xmm4, [esp + nb204_jyH1] subpd xmm5, [esp + nb204_jzH1] movapd [esp + nb204_dxMM], xmm0 movapd [esp + nb204_dyMM], xmm1 movapd [esp + nb204_dzMM], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb204_dxMH1], xmm3 movapd [esp + nb204_dyMH1], xmm4 movapd [esp + nb204_dzMH1], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 movapd [esp + nb204_rsqMM], xmm0 movapd [esp + nb204_rsqMH1], xmm3 movapd xmm0, [esp + nb204_ixM] movapd xmm1, [esp + nb204_iyM] movapd xmm2, [esp + nb204_izM] movapd xmm3, [esp + nb204_ixH1] movapd xmm4, [esp + nb204_iyH1] movapd xmm5, [esp + nb204_izH1] subpd xmm0, [esp + nb204_jxH2] subpd xmm1, [esp + nb204_jyH2] subpd xmm2, [esp + nb204_jzH2] subpd xmm3, [esp + nb204_jxM] subpd xmm4, [esp + nb204_jyM] subpd xmm5, [esp + nb204_jzM] movapd [esp + nb204_dxMH2], xmm0 movapd [esp + nb204_dyMH2], xmm1 movapd [esp + nb204_dzMH2], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb204_dxH1M], xmm3 movapd [esp + nb204_dyH1M], xmm4 movapd [esp + nb204_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 + nb204_rsqMH2], xmm0 movapd [esp + nb204_rsqH1M], xmm3 movapd xmm0, [esp + nb204_ixH1] movapd xmm1, [esp + nb204_iyH1] movapd xmm2, [esp + nb204_izH1] movapd xmm3, [esp + nb204_ixH1] movapd xmm4, [esp + nb204_iyH1] movapd xmm5, [esp + nb204_izH1] subpd xmm0, [esp + nb204_jxH1] subpd xmm1, [esp + nb204_jyH1] subpd xmm2, [esp + nb204_jzH1] subpd xmm3, [esp + nb204_jxH2] subpd xmm4, [esp + nb204_jyH2] subpd xmm5, [esp + nb204_jzH2] movapd [esp + nb204_dxH1H1], xmm0 movapd [esp + nb204_dyH1H1], xmm1 movapd [esp + nb204_dzH1H1], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb204_dxH1H2], xmm3 movapd [esp + nb204_dyH1H2], xmm4 movapd [esp + nb204_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 + nb204_rsqH1H1], xmm0 movapd [esp + nb204_rsqH1H2], xmm3 movapd xmm0, [esp + nb204_ixH2] movapd xmm1, [esp + nb204_iyH2] movapd xmm2, [esp + nb204_izH2] movapd xmm3, [esp + nb204_ixH2] movapd xmm4, [esp + nb204_iyH2] movapd xmm5, [esp + nb204_izH2] subpd xmm0, [esp + nb204_jxM] subpd xmm1, [esp + nb204_jyM] subpd xmm2, [esp + nb204_jzM] subpd xmm3, [esp + nb204_jxH1] subpd xmm4, [esp + nb204_jyH1] subpd xmm5, [esp + nb204_jzH1] movapd [esp + nb204_dxH2M], xmm0 movapd [esp + nb204_dyH2M], xmm1 movapd [esp + nb204_dzH2M], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [esp + nb204_dxH2H1], xmm3 movapd [esp + nb204_dyH2H1], xmm4 movapd [esp + nb204_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 + nb204_rsqH2M], xmm0 movapd [esp + nb204_rsqH2H1], xmm4 movapd xmm0, [esp + nb204_ixH2] movapd xmm1, [esp + nb204_iyH2] movapd xmm2, [esp + nb204_izH2] subpd xmm0, [esp + nb204_jxH2] subpd xmm1, [esp + nb204_jyH2] subpd xmm2, [esp + nb204_jzH2] movapd [esp + nb204_dxH2H2], xmm0 movapd [esp + nb204_dyH2H2], xmm1 movapd [esp + nb204_dzH2H2], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 addpd xmm0, xmm1 addpd xmm0, xmm2 movapd [esp + nb204_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 + nb204_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 + nb204_half] ;# iter1 mulpd xmm7, [esp + nb204_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 + nb204_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 + nb204_half] ;# rinv mulpd xmm5, [esp + nb204_half] ;# rinv movapd [esp + nb204_rinvH2H2], xmm1 movapd [esp + nb204_rinvH2H1], xmm5 movapd xmm0, [esp + nb204_rsqMM] movapd xmm4, [esp + nb204_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 + nb204_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 + nb204_half] ;# iter1 of mulpd xmm7, [esp + nb204_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 + nb204_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 + nb204_half] ;# rinv mulpd xmm5, [esp + nb204_half] ;# rinv movapd [esp + nb204_rinvMM], xmm1 movapd [esp + nb204_rinvMH1], xmm5 movapd xmm0, [esp + nb204_rsqMH2] movapd xmm4, [esp + nb204_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 + nb204_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 + nb204_half] ;# iter1 mulpd xmm7, [esp + nb204_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 + nb204_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 + nb204_half] ;# rinv mulpd xmm5, [esp + nb204_half] ;# rinv movapd [esp + nb204_rinvMH2], xmm1 movapd [esp + nb204_rinvH1M], xmm5 movapd xmm0, [esp + nb204_rsqH1H1] movapd xmm4, [esp + nb204_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 + nb204_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 + nb204_half] ;# iter1a mulpd xmm7, [esp + nb204_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 + nb204_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 + nb204_half] ;# rinv mulpd xmm5, [esp + nb204_half] ;# rinv movapd [esp + nb204_rinvH1H1], xmm1 movapd [esp + nb204_rinvH1H2], xmm5 movapd xmm0, [esp + nb204_rsqH2M] cvtpd2ps xmm1, xmm0 rsqrtps xmm1, xmm1 cvtps2pd xmm1, xmm1 movapd xmm2, xmm1 ;# copy of luA mulpd xmm1, xmm1 ;# luA*luA movapd xmm3, [esp + nb204_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 + nb204_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA mulpd xmm3, xmm3 ;# luA*luA movapd xmm1, [esp + nb204_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 + nb204_half] ;# rinv movapd [esp + nb204_rinvH2M], xmm1 ;# start with MM interaction movapd xmm0, [esp + nb204_rinvMM] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb204_krf] mulpd xmm0, xmm0 ;# rinvsq mulpd xmm5, [esp + nb204_rsqMM] ;# xmm5=krsq movapd xmm6, xmm5 addpd xmm6, xmm7 ;# xmm6=rinv+ krsq subpd xmm6, [esp + nb204_crf] mulpd xmm6, [esp + nb204_qqMM] ;# xmm6=voul=qq*(rinv+ krsq-crf) mulpd xmm5, [esp + nb204_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb204_qqMM] ;# xmm7 = coul part of fscal addpd xmm6, [esp + nb204_vctot] ;# local vctot summation variable mulpd xmm0, xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb204_dxMM] mulpd xmm1, [esp + nb204_dyMM] mulpd xmm2, [esp + nb204_dzMM] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb204_fixM] addpd xmm1, [esp + nb204_fiyM] addpd xmm2, [esp + nb204_fizM] movapd [esp + nb204_fjxM], xmm3 movapd [esp + nb204_fjyM], xmm4 movapd [esp + nb204_fjzM], xmm5 movapd [esp + nb204_fixM], xmm0 movapd [esp + nb204_fiyM], xmm1 movapd [esp + nb204_fizM], xmm2 ;# M-H1 interaction movapd xmm0, [esp + nb204_rinvMH1] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb204_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb204_rsqMH1] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=rinv+ krsq mulpd xmm0, xmm0 subpd xmm4, [esp + nb204_crf] mulpd xmm4, [esp + nb204_qqMH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb204_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb204_qqMH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsMH1 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb204_dxMH1] mulpd xmm1, [esp + nb204_dyMH1] mulpd xmm2, [esp + nb204_dzMH1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb204_fixM] addpd xmm1, [esp + nb204_fiyM] addpd xmm2, [esp + nb204_fizM] movapd [esp + nb204_fjxH1], xmm3 movapd [esp + nb204_fjyH1], xmm4 movapd [esp + nb204_fjzH1], xmm5 movapd [esp + nb204_fixM], xmm0 movapd [esp + nb204_fiyM], xmm1 movapd [esp + nb204_fizM], xmm2 ;# M-H2 interaction movapd xmm0, [esp + nb204_rinvMH2] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb204_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb204_rsqMH2] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq mulpd xmm0, xmm0 subpd xmm4, [esp + nb204_crf] mulpd xmm4, [esp + nb204_qqMH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb204_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb204_qqMH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb204_dxMH2]
⌨️ 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?