nb_kernel114_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,222 行 · 第 1/5 页
S
2,222 行
movsd [esp + nb114_vctot], xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb114_dxH1H2] mulsd xmm1, [esp + nb114_dyH1H2] mulsd xmm2, [esp + nb114_dzH1H2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb114_fixH1] addsd xmm1, [esp + nb114_fiyH1] addsd xmm2, [esp + nb114_fizH1] movsd [esp + nb114_fjxH2], xmm3 movsd [esp + nb114_fjyH2], xmm4 movsd [esp + nb114_fjzH2], xmm5 movsd [esp + nb114_fixH1], xmm0 movsd [esp + nb114_fiyH1], xmm1 movsd [esp + nb114_fizH1], xmm2 ;# H1-M interaction movsd xmm0, [esp + nb114_rinvH1M] movsd xmm7, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm7, [esp + nb114_qqMH] ;# vcoul mulsd xmm0, xmm7 addsd xmm7, [esp + nb114_vctot] movsd [esp + nb114_vctot], xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb114_dxH1M] mulsd xmm1, [esp + nb114_dyH1M] mulsd xmm2, [esp + nb114_dzH1M] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb114_fixH1] addsd xmm1, [esp + nb114_fiyH1] addsd xmm2, [esp + nb114_fizH1] movsd [esp + nb114_fjxM], xmm3 movsd [esp + nb114_fjyM], xmm4 movsd [esp + nb114_fjzM], xmm5 movsd [esp + nb114_fixH1], xmm0 movsd [esp + nb114_fiyH1], xmm1 movsd [esp + nb114_fizH1], xmm2 ;# H2-H1 interaction movsd xmm0, [esp + nb114_rinvH2H1] movsd xmm7, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm7, [esp + nb114_qqHH] ;# vcoul mulsd xmm0, xmm7 addsd xmm7, [esp + nb114_vctot] movsd [esp + nb114_vctot], xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb114_fjxH1] movapd xmm4, [esp + nb114_fjyH1] movapd xmm5, [esp + nb114_fjzH1] mulsd xmm0, [esp + nb114_dxH2H1] mulsd xmm1, [esp + nb114_dyH2H1] mulsd xmm2, [esp + nb114_dzH2H1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb114_fixH2] addsd xmm1, [esp + nb114_fiyH2] addsd xmm2, [esp + nb114_fizH2] movsd [esp + nb114_fjxH1], xmm3 movsd [esp + nb114_fjyH1], xmm4 movsd [esp + nb114_fjzH1], xmm5 movsd [esp + nb114_fixH2], xmm0 movsd [esp + nb114_fiyH2], xmm1 movsd [esp + nb114_fizH2], xmm2 ;# H2-H2 interaction movsd xmm0, [esp + nb114_rinvH2H2] movsd xmm7, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm7, [esp + nb114_qqHH] ;# vcoul mulsd xmm0, xmm7 addsd xmm7, [esp + nb114_vctot] movsd [esp + nb114_vctot], xmm7 movsd xmm1, xmm0 movsd xmm2, xmm0 movsd xmm3, [esp + nb114_fjxH2] movsd xmm4, [esp + nb114_fjyH2] movsd xmm5, [esp + nb114_fjzH2] mulsd xmm0, [esp + nb114_dxH2H2] mulsd xmm1, [esp + nb114_dyH2H2] mulsd xmm2, [esp + nb114_dzH2H2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb114_fixH2] addsd xmm1, [esp + nb114_fiyH2] addsd xmm2, [esp + nb114_fizH2] movsd [esp + nb114_fjxH2], xmm3 movsd [esp + nb114_fjyH2], xmm4 movsd [esp + nb114_fjzH2], xmm5 movsd [esp + nb114_fixH2], xmm0 movsd [esp + nb114_fiyH2], xmm1 movsd [esp + nb114_fizH2], xmm2 ;# H2-M interaction movsd xmm0, [esp + nb114_rinvH2M] movsd xmm7, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm7, [esp + nb114_qqMH] mulsd xmm0, xmm7 addsd xmm7, [esp + nb114_vctot] movsd [esp + nb114_vctot], xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb114_fjxM] movapd xmm4, [esp + nb114_fjyM] movapd xmm5, [esp + nb114_fjzM] mulsd xmm0, [esp + nb114_dxH2M] mulsd xmm1, [esp + nb114_dyH2M] mulsd xmm2, [esp + nb114_dzH2M] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb114_fixH2] addsd xmm1, [esp + nb114_fiyH2] addsd xmm2, [esp + nb114_fizH2] movsd [esp + nb114_fjxM], xmm3 movsd [esp + nb114_fjyM], xmm4 movsd [esp + nb114_fjzM], xmm5 movsd [esp + nb114_fixH2], xmm0 movsd [esp + nb114_fiyH2], xmm1 movsd [esp + nb114_fizH2], xmm2 ;# M-H1 interaction movsd xmm0, [esp + nb114_rinvMH1] movsd xmm7, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm7, [esp + nb114_qqMH] mulsd xmm0, xmm7 addsd xmm7, [esp + nb114_vctot] ;# local vctot summation variable movsd [esp + nb114_vctot], xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb114_fjxH1] movapd xmm4, [esp + nb114_fjyH1] movapd xmm5, [esp + nb114_fjzH1] mulsd xmm0, [esp + nb114_dxMH1] mulsd xmm1, [esp + nb114_dyMH1] mulsd xmm2, [esp + nb114_dzMH1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb114_fixM] addsd xmm1, [esp + nb114_fiyM] addsd xmm2, [esp + nb114_fizM] movsd [esp + nb114_fjxH1], xmm3 movsd [esp + nb114_fjyH1], xmm4 movsd [esp + nb114_fjzH1], xmm5 movsd [esp + nb114_fixM], xmm0 movsd [esp + nb114_fiyM], xmm1 movsd [esp + nb114_fizM], xmm2 ;# M-H2 interaction movsd xmm0, [esp + nb114_rinvMH2] movsd xmm7, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm7, [esp + nb114_qqMH] mulsd xmm0, xmm7 addsd xmm7, [esp + nb114_vctot] movsd [esp + nb114_vctot], xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb114_fjxH2] movapd xmm4, [esp + nb114_fjyH2] movapd xmm5, [esp + nb114_fjzH2] mulsd xmm0, [esp + nb114_dxMH2] mulsd xmm1, [esp + nb114_dyMH2] mulsd xmm2, [esp + nb114_dzMH2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb114_fixM] addsd xmm1, [esp + nb114_fiyM] addsd xmm2, [esp + nb114_fizM] movsd [esp + nb114_fjxH2], xmm3 movsd [esp + nb114_fjyH2], xmm4 movsd [esp + nb114_fjzH2], xmm5 movsd [esp + nb114_fixM], xmm0 movsd [esp + nb114_fiyM], xmm1 movsd [esp + nb114_fizM], xmm2 ;# M-M interaction movsd xmm0, [esp + nb114_rinvMM] movsd xmm7, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm7, [esp + nb114_qqMM] mulsd xmm0, xmm7 addsd xmm7, [esp + nb114_vctot] movsd [esp + nb114_vctot], xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb114_fjxM] movapd xmm4, [esp + nb114_fjyM] movapd xmm5, [esp + nb114_fjzM] mulsd xmm0, [esp + nb114_dxMM] mulsd xmm1, [esp + nb114_dyMM] mulsd xmm2, [esp + nb114_dzMM] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb114_fixM] addsd xmm1, [esp + nb114_fiyM] addsd xmm2, [esp + nb114_fizM] movsd [esp + nb114_fjxM], xmm3 movsd [esp + nb114_fjyM], xmm4 movsd [esp + nb114_fjzM], xmm5 movsd [esp + nb114_fixM], xmm0 movsd [esp + nb114_fiyM], xmm1 movsd [esp + nb114_fizM], xmm2 mov edi, [ebp + nb114_faction] ;# Did all interactions - now update j forces ;# Step1 - merge forces movlpd xmm0, [esp + nb114_fjxO] movlpd xmm1, [esp + nb114_fjzO] movlpd xmm2, [esp + nb114_fjyH1] movlpd xmm3, [esp + nb114_fjxH2] movlpd xmm4, [esp + nb114_fjzH2] movlpd xmm5, [esp + nb114_fjyM] movhpd xmm0, [esp + nb114_fjyO] movhpd xmm1, [esp + nb114_fjxH1] movhpd xmm2, [esp + nb114_fjzH1] movhpd xmm3, [esp + nb114_fjyH2] movhpd xmm4, [esp + nb114_fjxM] movhpd xmm5, [esp + nb114_fjzM] movlps xmm6, [edi + eax*8] movhps xmm6, [edi + eax*8 + 8] movlps xmm7, [edi + eax*8 + 16] movhps xmm7, [edi + eax*8 + 24] addpd xmm0, xmm6 addpd xmm1, xmm7 movlps xmm6, [edi + eax*8 + 32] movhps xmm6, [edi + eax*8 + 40] movlps xmm7, [edi + eax*8 + 48] movhps xmm7, [edi + eax*8 + 56] addpd xmm2, xmm6 addpd xmm3, xmm7 movlps xmm6, [edi + eax*8 + 64] movhps xmm6, [edi + eax*8 + 72] movlps xmm7, [edi + eax*8 + 80] movhps xmm7, [edi + eax*8 + 88] addpd xmm4, xmm6 addpd xmm5, xmm7 movlps [edi + eax*8], xmm0 movhps [edi + eax*8 + 8], xmm0 movlps [edi + eax*8 + 16], xmm1 movhps [edi + eax*8 + 24], xmm1 movlps [edi + eax*8 + 32], xmm2 movhps [edi + eax*8 + 40], xmm2 movlps [edi + eax*8 + 48], xmm3 movhps [edi + eax*8 + 56], xmm3 movlps [edi + eax*8 + 64], xmm4 movhps [edi + eax*8 + 72], xmm4 movlps [edi + eax*8 + 80], xmm5 movhps [edi + eax*8 + 88], xmm5 .nb114_updateouterdata: mov ecx, [esp + nb114_ii3] mov edi, [ebp + nb114_faction] mov esi, [ebp + nb114_fshift] mov edx, [esp + nb114_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movapd xmm0, [esp + nb114_fixO] movapd xmm1, [esp + nb114_fiyO] movapd xmm2, [esp + nb114_fizO] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addsd xmm0, xmm3 addsd xmm1, xmm4 addsd xmm2, xmm5 ;# sum is in low xmm0-xmm2 movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 ;# increment i force movsd xmm3, [edi + ecx*8] movsd xmm4, [edi + ecx*8 + 8] movsd xmm5, [edi + ecx*8 + 16] addsd xmm3, xmm0 addsd xmm4, xmm1 addsd xmm5, xmm2 movsd [edi + ecx*8], xmm3 movsd [edi + ecx*8 + 8], xmm4 movsd [edi + ecx*8 + 16], xmm5 ;# accumulate force in xmm6/xmm7 for fshift movapd xmm6, xmm0 movsd xmm7, xmm2 unpcklpd xmm6, xmm1 ;# accumulate H1i forces in xmm0, xmm1, xmm2 movapd xmm0, [esp + nb114_fixH1] movapd xmm1, [esp + nb114_fiyH1] movapd xmm2, [esp + nb114_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 addsd xmm7, xmm2 unpcklpd xmm0, xmm1 addpd xmm6, xmm0 ;# accumulate H2i forces in xmm0, xmm1, xmm2 movapd xmm0, [esp + nb114_fixH2] movapd xmm1, [esp + nb114_fiyH2] movapd xmm2, [esp + nb114_fizH2] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addsd xmm0, xmm3 addsd xmm1, xmm4 addsd xmm2, xmm5 ;# sum is in low xmm0-xmm2 movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 ;# increment i force movsd xmm3, [edi + ecx*8 + 48] movsd xmm4, [edi + ecx*8 + 56] movsd xmm5, [edi + ecx*8 + 64] addsd xmm3, xmm0 addsd xmm4, xmm1 addsd xmm5, xmm2 movsd [edi + ecx*8 + 48], xmm3 movsd [edi + ecx*8 + 56], xmm4 movsd [edi + ecx*8 + 64], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addsd xmm7, xmm2 unpcklpd xmm0, xmm1 addpd xmm6, xmm0 ;# accumulate Mi forces in xmm0, xmm1, xmm2 movapd xmm0, [esp + nb114_fixM] movapd xmm1, [esp + nb114_fiyM] movapd xmm2, [esp + nb114_fizM] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addsd xmm0, xmm3 addsd xmm1, xmm4 addsd xmm2, xmm5 ;# sum is in low xmm0-xmm2 movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 ;# increment i force movsd xmm3, [edi + ecx*8 + 72] movsd xmm4, [edi + ecx*8 + 80] movsd xmm5, [edi + ecx*8 + 88] addsd xmm3, xmm0 addsd xmm4, xmm1 addsd xmm5, xmm2 movsd [edi + ecx*8 + 72], xmm3 movsd [edi + ecx*8 + 80], xmm4 movsd [edi + ecx*8 + 88], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addsd xmm7, xmm2 unpcklpd xmm0, xmm1 addpd xmm6, xmm0 ;# increment fshift force movlpd xmm3, [esi + edx*8] movhpd xmm3, [esi + edx*8 + 8] movsd xmm4, [esi + edx*8 + 16] addpd xmm3, xmm6 addsd xmm4, xmm7 movlpd [esi + edx*8], xmm3 movhpd [esi + edx*8 + 8], xmm3 movsd [esi + edx*8 + 16], xmm4 ;# get n from stack mov esi, [esp + nb11
⌨️ 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?