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
mulsd 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] 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 + nb104_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 + nb104_half] ;# iter1 mulsd xmm7, [esp + nb104_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 + nb104_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 + nb104_half] ;# rinv mulsd 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] 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 + nb104_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 + nb104_half] ;# iter1a mulsd xmm7, [esp + nb104_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 + nb104_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 + nb104_half] ;# rinv mulsd xmm5, [esp + nb104_half] ;# rinv movapd [esp + nb104_rinvH1H1], xmm1 movapd [esp + nb104_rinvH1H2], xmm5 movapd xmm0, [esp + nb104_rsqH2M] cvtsd2ss xmm1, xmm0 rsqrtss xmm1, xmm1 cvtss2sd xmm1, xmm1 movapd xmm2, xmm1 ;# copy of luA mulsd xmm1, xmm1 ;# luA*luA movapd xmm3, [esp + nb104_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 + nb104_half] ;# iter1 movapd xmm2, xmm3 ;# copy of luA mulsd xmm3, xmm3 ;# luA*luA movapd xmm1, [esp + nb104_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 + nb104_half] ;# rinv movapd [esp + nb104_rinvH2M], xmm1 ;# start with MM interaction movapd xmm0, [esp + nb104_rinvMM] movapd xmm7, xmm0 mulsd xmm0, xmm0 mulsd xmm7, [esp + nb104_qqMM] mulsd xmm0, xmm7 addsd xmm7, [esp + nb104_vctot] movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb104_dxMM] mulsd xmm1, [esp + nb104_dyMM] mulsd xmm2, [esp + nb104_dzMM] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb104_fixM] addsd xmm1, [esp + nb104_fiyM] addsd xmm2, [esp + nb104_fizM] movlpd [esp + nb104_fjxM], xmm3 movlpd [esp + nb104_fjyM], xmm4 movlpd [esp + nb104_fjzM], xmm5 movlpd [esp + nb104_fixM], xmm0 movlpd [esp + nb104_fiyM], xmm1 movlpd [esp + nb104_fizM], xmm2 ;# M-H1 interaction movapd xmm0, [esp + nb104_rinvMH1] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb104_qqMH] mulsd xmm0, xmm1 ;# fsMH1 addsd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb104_dxMH1] mulsd xmm1, [esp + nb104_dyMH1] mulsd xmm2, [esp + nb104_dzMH1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb104_fixM] addsd xmm1, [esp + nb104_fiyM] addsd xmm2, [esp + nb104_fizM] movlpd [esp + nb104_fjxH1], xmm3 movlpd [esp + nb104_fjyH1], xmm4 movlpd [esp + nb104_fjzH1], xmm5 movlpd [esp + nb104_fixM], xmm0 movlpd [esp + nb104_fiyM], xmm1 movlpd [esp + nb104_fizM], xmm2 ;# M-H2 interaction movapd xmm0, [esp + nb104_rinvMH2] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb104_qqMH] mulsd xmm0, xmm1 ;# fsMH2 addsd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb104_dxMH2] mulsd xmm1, [esp + nb104_dyMH2] mulsd xmm2, [esp + nb104_dzMH2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb104_fixM] addsd xmm1, [esp + nb104_fiyM] addsd xmm2, [esp + nb104_fizM] movlpd [esp + nb104_fjxH2], xmm3 movlpd [esp + nb104_fjyH2], xmm4 movlpd [esp + nb104_fjzH2], xmm5 movlpd [esp + nb104_fixM], xmm0 movlpd [esp + nb104_fiyM], xmm1 movlpd [esp + nb104_fizM], xmm2 ;# H1-M interaction movapd xmm0, [esp + nb104_rinvH1M] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb104_qqMH] mulsd xmm0, xmm1 ;# fsH1M addsd 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] mulsd xmm0, [esp + nb104_dxH1M] mulsd xmm1, [esp + nb104_dyH1M] mulsd xmm2, [esp + nb104_dzH1M] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb104_fixH1] addsd xmm1, [esp + nb104_fiyH1] addsd xmm2, [esp + nb104_fizH1] movlpd [esp + nb104_fjxM], xmm3 movlpd [esp + nb104_fjyM], xmm4 movlpd [esp + nb104_fjzM], xmm5 movlpd [esp + nb104_fixH1], xmm0 movlpd [esp + nb104_fiyH1], xmm1 movlpd [esp + nb104_fizH1], xmm2 ;# H1-H1 interaction movapd xmm0, [esp + nb104_rinvH1H1] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb104_qqHH] mulsd xmm0, xmm1 ;# fsH1H1 addsd 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] mulsd xmm0, [esp + nb104_dxH1H1] mulsd xmm1, [esp + nb104_dyH1H1] mulsd xmm2, [esp + nb104_dzH1H1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb104_fixH1] addsd xmm1, [esp + nb104_fiyH1] addsd xmm2, [esp + nb104_fizH1] movlpd [esp + nb104_fjxH1], xmm3 movlpd [esp + nb104_fjyH1], xmm4 movlpd [esp + nb104_fjzH1], xmm5 movlpd [esp + nb104_fixH1], xmm0 movlpd [esp + nb104_fiyH1], xmm1 movlpd [esp + nb104_fizH1], xmm2 ;# H1-H2 interaction movapd xmm0, [esp + nb104_rinvH1H2] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb104_qqHH] mulsd xmm0, xmm1 ;# fsMH2 addsd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb104_fjxH2] movapd xmm4, [esp + nb104_fjyH2] movapd xmm5, [esp + nb104_fjzH2] mulsd xmm0, [esp + nb104_dxH1H2] mulsd xmm1, [esp + nb104_dyH1H2] mulsd xmm2, [esp + nb104_dzH1H2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb104_fixH1] addsd xmm1, [esp + nb104_fiyH1] addsd xmm2, [esp + nb104_fizH1] movlpd [esp + nb104_fjxH2], xmm3 movlpd [esp + nb104_fjyH2], xmm4 movlpd [esp + nb104_fjzH2], xmm5 movlpd [esp + nb104_fixH1], xmm0 movlpd [esp + nb104_fiyH1], xmm1 movlpd [esp + nb104_fizH1], xmm2 ;# H2-M interaction movapd xmm0, [esp + nb104_rinvH2M] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb104_qqMH] mulsd xmm0, xmm1 ;# fsH2M addsd 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] mulsd xmm0, [esp + nb104_dxH2M] mulsd xmm1, [esp + nb104_dyH2M] mulsd xmm2, [esp + nb104_dzH2M] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb104_fixH2] addsd xmm1, [esp + nb104_fiyH2] addsd xmm2, [esp + nb104_fizH2] movlpd [esp + nb104_fjxM], xmm3 movlpd [esp + nb104_fjyM], xmm4 movlpd [esp + nb104_fjzM], xmm5 movlpd [esp + nb104_fixH2], xmm0 movlpd [esp + nb104_fiyH2], xmm1 movlpd [esp + nb104_fizH2], xmm2 ;# H2-H1 interaction movapd xmm0, [esp + nb104_rinvH2H1] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb104_qqHH] mulsd xmm0, xmm1 ;# fsH2H1 addsd 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] mulsd xmm0, [esp + nb104_dxH2H1] mulsd xmm1, [esp + nb104_dyH2H1] mulsd xmm2, [esp + nb104_dzH2H1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb104_fixH2] addsd xmm1, [esp + nb104_fiyH2] addsd xmm2, [esp + nb104_fizH2] movlpd [esp + nb104_fjxH1], xmm3 movlpd [esp + nb104_fjyH1], xmm4 movlpd [esp + nb104_fjzH1], xmm5 movlpd [esp + nb104_fixH2], xmm0 movlpd [esp + nb104_fiyH2], xmm1 movlpd [esp + nb104_fizH2], xmm2 ;# H2-H2 interaction movapd xmm0, [esp + nb104_rinvH2H2] movapd xmm1, xmm0 mulsd xmm0, xmm0 mulsd xmm1, [esp + nb104_qqHH] mulsd xmm0, xmm1 ;# fsH2H2 addsd xmm7, xmm1 ;# add to local vctot movapd xmm1, xmm0 movsd [esp + nb104_vctot], xmm7 movapd xmm2, xmm0 movapd xmm3, [esp + nb104_fjxH2] movapd xmm4, [esp + nb104_fjyH2] movapd xmm5, [esp + nb104_fjzH2] mulsd xmm0, [esp + nb104_dxH2H2] mulsd xmm1, [esp + nb104_dyH2H2] mulsd xmm2, [esp + nb104_dzH2H2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb104_fixH2] addsd xmm1, [esp + nb104_fiyH2] addsd xmm2, [esp + nb104_fizH2] movlpd [esp + nb104_fjxH2], xmm3 movlpd [esp + nb104_fjyH2], xmm4 movlpd [esp + nb104_fjzH2], xmm5 movlpd [esp + nb104_fixH2], xmm0 movlpd [esp + nb104_fiyH2], xmm1 movlpd [esp + nb104_fizH2], xmm2 mov edi, [ebp + nb104_faction] ;# Did all interactions - now update j forces movlpd xmm0, [edi + eax*8 + 24] movlpd xmm1, [edi + eax*8 + 32] movlpd xmm2, [edi + eax*8 + 40] movlpd xmm3, [edi + eax*8 + 48] movlpd xmm4, [edi + eax*8 + 56] movlpd xmm5, [edi + eax*8 + 64] movlpd xmm6, [edi + eax*8 + 72] movlpd xmm7, [edi + eax*8 + 80] addsd xmm0, [esp + nb104_fjxH1] addsd xmm1, [esp + nb104_fjyH1] addsd xmm2, [esp + nb104_fjzH1] addsd xmm3, [esp + nb104_fjxH2] addsd xmm4, [esp + nb104_fjyH2] addsd xmm5, [esp + nb104_fjzH2] addsd xmm6, [esp + nb104_fjxM] addsd xmm7, [esp + nb104_fjyM] movlpd [edi + eax*8 + 24], xmm0 movlpd [edi + eax*8 + 32], xmm1 movlpd [edi + eax*8 + 40], xmm2 movlpd [edi + eax*8 + 48], xmm3 movlpd [edi + eax*8 + 56], xmm4 movlpd [edi + eax*8 + 64], xmm5 movlpd [edi + eax*8 + 72], xmm6 movlpd [edi + eax*8 + 80], xmm7 movlpd xmm0, [edi + eax*8 + 88] addsd xmm0, [esp + nb104_fjzM] movlpd [edi + eax*8 + 88], xmm0 .nb104_updateouterdata: mov ecx, [esp + nb104_ii3] mov edi, [ebp + nb104_faction] mov esi, [ebp + nb104_fshift] mov edx, [esp + nb104_is3] ;# accumulate H1i forces in xmm0, xmm1, xmm2 movapd xmm0, [esp + nb104_fixH1] movapd xmm1, [esp + nb104_fiyH1] movapd xmm2, [esp + nb104_fizH1] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addsd xmm0, xmm3 addsd xmm1, xmm4 addsd xmm2, xmm5 ;# sum is in low xmm0-xmm2 ;# increment i force movsd xmm3, [edi + ecx*8 + 24] movsd xmm4, [edi + ecx*8 + 32] movsd xmm5, [edi + ecx*8 + 40] addsd xmm3, xmm0 addsd xmm4, xmm1 addsd xmm5, xmm2 movsd [edi + ecx*8 + 24], xmm3 movsd [edi + ecx*8 + 32], xmm4 movsd [edi + ecx*8 + 40], xmm5 ;# accumulate force in xmm6/xmm7 for fshift movapd xmm6, xmm0 movsd xmm7, xmm2 unpcklpd xmm6, xmm1 ;# accumulate H2i forces in xmm0, xmm1, 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?