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 [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] subsd xmm0, [esp + nb204_jxM] subsd xmm1, [esp + nb204_jyM] subsd xmm2, [esp + nb204_jzM] subsd xmm3, [esp + nb204_jxH1] subsd xmm4, [esp + nb204_jyH1] subsd xmm5, [esp + nb204_jzH1] movapd [esp + nb204_dxH2M], xmm0 movapd [esp + nb204_dyH2M], xmm1 movapd [esp + nb204_dzH2M], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb204_dxH2H1], xmm3 movapd [esp + nb204_dyH2H1], xmm4 movapd [esp + nb204_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 + nb204_rsqH2M], xmm0 movapd [esp + nb204_rsqH2H1], xmm4 movapd xmm0, [esp + nb204_ixH2] movapd xmm1, [esp + nb204_iyH2] movapd xmm2, [esp + nb204_izH2] subsd xmm0, [esp + nb204_jxH2] subsd xmm1, [esp + nb204_jyH2] subsd xmm2, [esp + nb204_jzH2] movapd [esp + nb204_dxH2H2], xmm0 movapd [esp + nb204_dyH2H2], xmm1 movapd [esp + nb204_dzH2H2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 addsd xmm0, xmm1 addsd xmm0, xmm2 movapd [esp + nb204_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 + nb204_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 + nb204_half] ;# iter1 mulsd xmm7, [esp + nb204_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 + nb204_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 + nb204_half] ;# rinv mulsd 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] 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 + nb204_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 + nb204_half] ;# iter1 of mulsd xmm7, [esp + nb204_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 + nb204_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 + nb204_half] ;# rinv mulsd 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] 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 + nb204_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 + nb204_half] ;# iter1 mulsd xmm7, [esp + nb204_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 + nb204_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 + nb204_half] ;# rinv mulsd 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] 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 + nb204_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 + nb204_half] ;# iter1a mulsd xmm7, [esp + nb204_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 + nb204_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 + nb204_half] ;# rinv mulsd xmm5, [esp + nb204_half] ;# rinv movapd [esp + nb204_rinvH1H1], xmm1 movapd [esp + nb204_rinvH1H2], xmm5 movapd xmm0, [esp + nb204_rsqH2M] cvtsd2ss xmm1, xmm0 rsqrtss xmm1, xmm1 cvtss2sd xmm1, xmm1 movapd xmm2, xmm1 ;# copy of luA mulsd xmm1, xmm1 ;# luA*luA movapd xmm3, [esp + nb204_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 + nb204_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA mulsd xmm3, xmm3 ;# luA*luA movapd xmm1, [esp + nb204_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 + 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] mulsd xmm0, xmm0 movapd xmm1, xmm0 mulsd xmm1, xmm0 mulsd xmm1, xmm0 ;# xmm1=rinvsix mulsd xmm5, [esp + nb204_rsqMM] ;# xmm5=krsq movapd xmm6, xmm5 addsd xmm6, xmm7 ;# xmm6=rinv+ krsq subsd xmm6, [esp + nb204_crf] mulsd xmm6, [esp + nb204_qqMM] ;# xmm6=voul=qq*(rinv+ krsq-crf) mulsd xmm5, [esp + nb204_two] subsd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulsd xmm7, [esp + nb204_qqMM] ;# xmm7 = coul part of fscal addsd xmm6, [esp + nb204_vctot] ;# local vctot summation variable mulsd xmm0, xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb204_dxMM] mulsd xmm1, [esp + nb204_dyMM] mulsd xmm2, [esp + nb204_dzMM] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb204_fixM] addsd xmm1, [esp + nb204_fiyM] addsd xmm2, [esp + nb204_fizM] movlpd [esp + nb204_fjxM], xmm3 movlpd [esp + nb204_fjyM], xmm4 movlpd [esp + nb204_fjzM], xmm5 movlpd [esp + nb204_fixM], xmm0 movlpd [esp + nb204_fiyM], xmm1 movlpd [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 mulsd xmm5, [esp + nb204_rsqMH1] ;# xmm5=krsq movapd xmm4, xmm5 addsd xmm4, xmm7 ;# xmm4=rinv+ krsq mulsd xmm0, xmm0 subsd xmm4, [esp + nb204_crf] mulsd xmm4, [esp + nb204_qqMH] ;# xmm4=voul=qq*(rinv+ krsq) mulsd xmm5, [esp + nb204_two] subsd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulsd xmm7, [esp + nb204_qqMH] ;# xmm7 = coul part of fscal addsd xmm6, xmm4 ;# add to local vctot mulsd xmm0, xmm7 ;# fsMH1 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb204_dxMH1] mulsd xmm1, [esp + nb204_dyMH1] mulsd xmm2, [esp + nb204_dzMH1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb204_fixM] addsd xmm1, [esp + nb204_fiyM] addsd xmm2, [esp + nb204_fizM] movlpd [esp + nb204_fjxH1], xmm3 movlpd [esp + nb204_fjyH1], xmm4 movlpd [esp + nb204_fjzH1], xmm5 movlpd [esp + nb204_fixM], xmm0 movlpd [esp + nb204_fiyM], xmm1 movlpd [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 mulsd xmm5, [esp + nb204_rsqMH2] ;# xmm5=krsq movapd xmm4, xmm5 addsd xmm4, xmm7 ;# xmm4=r inv+ krsq mulsd xmm0, xmm0 subsd xmm4, [esp + nb204_crf] mulsd xmm4, [esp + nb204_qqMH] ;# xmm4=voul=qq*(rinv+ krsq) mulsd xmm5, [esp + nb204_two] subsd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulsd xmm7, [esp + nb204_qqMH] ;# xmm7 = coul part of fscal addsd xmm6, xmm4 ;# add to local vctot mulsd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb204_dxMH2] mulsd xmm1, [esp + nb204_dyMH2] mulsd xmm2, [esp + nb204_dzMH2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb204_fixM] addsd xmm1, [esp + nb204_fiyM] addsd xmm2, [esp + nb204_fizM] movlpd [esp + nb204_fjxH2], xmm3 movlpd [esp + nb204_fjyH2], xmm4 movlpd [esp + nb204_fjzH2], xmm5 movlpd [esp + nb204_fixM], xmm0 movlpd [esp + nb204_fiyM], xmm1 movlpd [esp + nb204_fizM], xmm2 ;# H1-M interaction movapd xmm0, [esp + nb204_rinvH1M] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb204_krf] movapd xmm1, xmm0 mulsd xmm5, [esp + nb204_rsqH1M] ;# xmm5=krsq movapd xmm4, xmm5 addsd xmm4, xmm7 ;# xmm4=rinv+ krsq mulsd xmm0, xmm0 subsd xmm4, [esp + nb204_crf] mulsd xmm4, [esp + nb204_qqMH] ;# xmm4=voul=qq*(rinv+ krsq) mulsd xmm5, [esp + nb204_two] subsd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulsd xmm7, [esp + nb204_qqMH] ;# xmm7 = coul part of fscal addsd xmm6, xmm4 ;# add to local vctot mulsd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxM] movapd xmm4, [esp + nb204_fjyM] movapd xmm5, [esp + nb204_fjzM] mulsd xmm0, [esp + nb204_dxH1M] mulsd xmm1, [esp + nb204_dyH1M] mulsd xmm2, [esp + nb204_dzH1M] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb204_fixH1] addsd xmm1, [esp + nb204_fiyH1] addsd xmm2, [esp + nb204_fizH1] movlpd [esp + nb204_fjxM], xmm3 movlpd [esp + nb204_fjyM], xmm4 movlpd [esp + nb204_fjzM], xmm5 movlpd [esp + nb204_fixH1], xmm0 movlpd [esp + nb204_fiyH1], xmm1 movlpd [esp + nb204_fizH1], xmm2 ;# H1-H1 interaction movapd xmm0, [esp + nb204_rinvH1H1] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb204_krf] movapd xmm1, xmm0 mulsd xmm5, [esp + nb204_rsqH1H1] ;# xmm5=krsq movapd xmm4, xmm5 addsd xmm4, xmm7 ;# xmm4=r inv+ krsq subsd xmm4, [esp + nb204_crf] mulsd xmm0, xmm0 mulsd xmm4, [esp + nb204_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulsd xmm5, [esp + nb204_two] subsd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulsd xmm7, [esp + nb204_qqHH] ;# xmm7 = coul part of fscal addsd xmm6, xmm4 ;# add to local vctot mulsd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxH1] movapd xmm4, [esp + nb204_fjyH1] movapd xmm5, [esp + nb204_fjzH1] mulsd xmm0, [esp + nb204_dxH1H1] mulsd xmm1, [esp + nb204_dyH1H1] mulsd xmm2, [esp + nb204_dzH1H1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb204_fixH1] addsd xmm1, [esp + nb204_fiyH1] addsd xmm2, [esp + nb204_fizH1] movlpd [esp + nb204_fjxH1], xmm3 movlpd [esp + nb204_fjyH1], xmm4⌨️ 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?