nb_kernel134_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,244 行 · 第 1/5 页
S
2,244 行
shl ebx, 3 mov esi, [ebp + nb134_VFtab] ;# dispersion movlpd xmm4, [esi + ebx*8] ;# Y1 movhpd xmm4, [esi + ebx*8 + 8] ;# Y1 F1 movapd xmm5, xmm4 unpcklpd xmm4, xmm3 ;# Y1 Y2 unpckhpd xmm5, xmm3 ;# F1 F2 movlpd xmm6, [esi + ebx*8 + 16] ;# G1 movhpd xmm6, [esi + ebx*8 + 24] ;# G1 H1 movapd xmm7, xmm6 unpcklpd xmm6, xmm3 ;# G1 G2 unpckhpd xmm7, xmm3 ;# H1 H2 ;# dispersion table ready, in xmm4-xmm7 mulsd xmm6, xmm1 ;# xmm6=Geps mulsd xmm7, xmm2 ;# xmm7=Heps2 addsd xmm5, xmm6 addsd xmm5, xmm7 ;# xmm5=Fp mulsd xmm7, [esp + nb134_two] ;# two*Heps2 addsd xmm7, xmm6 addsd xmm7, xmm5 ;# xmm7=FF mulsd xmm5, xmm1 ;# xmm5=eps*Fp addsd xmm5, xmm4 ;# xmm5=VV movsd xmm4, [esp + nb134_c6] mulsd xmm7, xmm4 ;# fijD mulsd xmm5, xmm4 ;# Vvdw6 ;# put scalar force on stack Update Vvdwtot directly addsd xmm5, [esp + nb134_Vvdwtot] xorpd xmm3, xmm3 mulsd xmm7, [esp + nb134_tsc] subsd xmm3, xmm7 movsd [esp + nb134_fstmp], xmm3 movsd [esp + nb134_Vvdwtot], xmm5 ;# repulsion movlpd xmm4, [esi + ebx*8 + 32] ;# Y1 movhpd xmm4, [esi + ebx*8 + 40] ;# Y1 F1 movapd xmm5, xmm4 unpcklpd xmm4, xmm3 ;# Y1 Y2 unpckhpd xmm5, xmm3 ;# F1 F2 movlpd xmm6, [esi + ebx*8 + 48] ;# G1 movhpd xmm6, [esi + ebx*8 + 56] ;# G1 H1 movapd xmm7, xmm6 unpcklpd xmm6, xmm3 ;# G1 G2 unpckhpd xmm7, xmm3 ;# H1 H2 ;# table ready, in xmm4-xmm7 mulsd xmm6, xmm1 ;# xmm6=Geps mulsd xmm7, xmm2 ;# xmm7=Heps2 addsd xmm5, xmm6 addsd xmm5, xmm7 ;# xmm5=Fp mulsd xmm7, [esp + nb134_two] ;# two*Heps2 addsd xmm7, xmm6 addsd xmm7, xmm5 ;# xmm7=FF mulsd xmm5, xmm1 ;# xmm5=eps*Fp addsd xmm5, xmm4 ;# xmm5=VV movsd xmm4, [esp + nb134_c12] mulsd xmm7, xmm4 mulsd xmm5, xmm4 addsd xmm5, [esp + nb134_Vvdwtot] movsd xmm3, [esp + nb134_fstmp] mulsd xmm7, [esp + nb134_tsc] subsd xmm3, xmm7 movsd [esp + nb134_Vvdwtot], xmm5 mulsd xmm0, xmm3 movsd xmm1, xmm0 movsd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb134_dxOO] mulsd xmm1, [esp + nb134_dyOO] mulsd xmm2, [esp + nb134_dzOO] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb134_fixO] addsd xmm1, [esp + nb134_fiyO] addsd xmm2, [esp + nb134_fizO] movsd [esp + nb134_fjxO], xmm3 movsd [esp + nb134_fjyO], xmm4 movsd [esp + nb134_fjzO], xmm5 movsd [esp + nb134_fixO], xmm0 movsd [esp + nb134_fiyO], xmm1 movsd [esp + nb134_fizO], xmm2 ;# H1-H1 interaction movsd xmm0, [esp + nb134_rinvH1H1] movsd xmm6, xmm0 ;# xmm6=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm6, [esp + nb134_qqHH] ;# vcoul mulsd xmm0, xmm6 addsd xmm6, [esp + nb134_vctot] movsd xmm1, xmm0 movsd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb134_dxH1H1] mulsd xmm1, [esp + nb134_dyH1H1] mulsd xmm2, [esp + nb134_dzH1H1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb134_fixH1] addsd xmm1, [esp + nb134_fiyH1] addsd xmm2, [esp + nb134_fizH1] movsd [esp + nb134_fjxH1], xmm3 movsd [esp + nb134_fjyH1], xmm4 movsd [esp + nb134_fjzH1], xmm5 movsd [esp + nb134_fixH1], xmm0 movsd [esp + nb134_fiyH1], xmm1 movsd [esp + nb134_fizH1], xmm2 ;# H1-H2 interaction movsd xmm0, [esp + nb134_rinvH1H2] movsd xmm4, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm4, [esp + nb134_qqHH] ;# vcoul mulsd xmm0, xmm4 addsd xmm6, xmm4 movsd xmm1, xmm0 movsd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb134_dxH1H2] mulsd xmm1, [esp + nb134_dyH1H2] mulsd xmm2, [esp + nb134_dzH1H2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb134_fixH1] addsd xmm1, [esp + nb134_fiyH1] addsd xmm2, [esp + nb134_fizH1] movsd [esp + nb134_fjxH2], xmm3 movsd [esp + nb134_fjyH2], xmm4 movsd [esp + nb134_fjzH2], xmm5 movsd [esp + nb134_fixH1], xmm0 movsd [esp + nb134_fiyH1], xmm1 movsd [esp + nb134_fizH1], xmm2 ;# H1-M interaction movsd xmm0, [esp + nb134_rinvH1M] movsd xmm4, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm4, [esp + nb134_qqMH] ;# vcoul mulsd xmm0, xmm4 addsd xmm6, xmm4 movsd xmm1, xmm0 movsd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulsd xmm0, [esp + nb134_dxH1M] mulsd xmm1, [esp + nb134_dyH1M] mulsd xmm2, [esp + nb134_dzH1M] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb134_fixH1] addsd xmm1, [esp + nb134_fiyH1] addsd xmm2, [esp + nb134_fizH1] movsd [esp + nb134_fjxM], xmm3 movsd [esp + nb134_fjyM], xmm4 movsd [esp + nb134_fjzM], xmm5 movsd [esp + nb134_fixH1], xmm0 movsd [esp + nb134_fiyH1], xmm1 movsd [esp + nb134_fizH1], xmm2 ;# H2-H1 interaction movsd xmm0, [esp + nb134_rinvH2H1] movsd xmm4, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm4, [esp + nb134_qqHH] ;# vcoul mulsd xmm0, xmm4 addsd xmm6, xmm4 movsd xmm1, xmm0 movsd xmm2, xmm0 movapd xmm3, [esp + nb134_fjxH1] movapd xmm4, [esp + nb134_fjyH1] movapd xmm5, [esp + nb134_fjzH1] mulsd xmm0, [esp + nb134_dxH2H1] mulsd xmm1, [esp + nb134_dyH2H1] mulsd xmm2, [esp + nb134_dzH2H1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb134_fixH2] addsd xmm1, [esp + nb134_fiyH2] addsd xmm2, [esp + nb134_fizH2] movsd [esp + nb134_fjxH1], xmm3 movsd [esp + nb134_fjyH1], xmm4 movsd [esp + nb134_fjzH1], xmm5 movsd [esp + nb134_fixH2], xmm0 movsd [esp + nb134_fiyH2], xmm1 movsd [esp + nb134_fizH2], xmm2 ;# H2-H2 interaction movsd xmm0, [esp + nb134_rinvH2H2] movsd xmm4, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm4, [esp + nb134_qqHH] ;# vcoul mulsd xmm0, xmm4 addsd xmm6, xmm4 movsd xmm1, xmm0 movsd xmm2, xmm0 movsd xmm3, [esp + nb134_fjxH2] movsd xmm4, [esp + nb134_fjyH2] movsd xmm5, [esp + nb134_fjzH2] mulsd xmm0, [esp + nb134_dxH2H2] mulsd xmm1, [esp + nb134_dyH2H2] mulsd xmm2, [esp + nb134_dzH2H2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb134_fixH2] addsd xmm1, [esp + nb134_fiyH2] addsd xmm2, [esp + nb134_fizH2] movsd [esp + nb134_fjxH2], xmm3 movsd [esp + nb134_fjyH2], xmm4 movsd [esp + nb134_fjzH2], xmm5 movsd [esp + nb134_fixH2], xmm0 movsd [esp + nb134_fiyH2], xmm1 movsd [esp + nb134_fizH2], xmm2 ;# H2-M interaction movsd xmm0, [esp + nb134_rinvH2M] movsd xmm4, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm4, [esp + nb134_qqMH] ;# vcoul mulsd xmm0, xmm4 addsd xmm6, xmm4 movsd xmm1, xmm0 movsd xmm2, xmm0 movapd xmm3, [esp + nb134_fjxM] movapd xmm4, [esp + nb134_fjyM] movapd xmm5, [esp + nb134_fjzM] mulsd xmm0, [esp + nb134_dxH2M] mulsd xmm1, [esp + nb134_dyH2M] mulsd xmm2, [esp + nb134_dzH2M] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb134_fixH2] addsd xmm1, [esp + nb134_fiyH2] addsd xmm2, [esp + nb134_fizH2] movsd [esp + nb134_fjxM], xmm3 movsd [esp + nb134_fjyM], xmm4 movsd [esp + nb134_fjzM], xmm5 movsd [esp + nb134_fixH2], xmm0 movsd [esp + nb134_fiyH2], xmm1 movsd [esp + nb134_fizH2], xmm2 ;# M-H1 interaction movsd xmm0, [esp + nb134_rinvMH1] movsd xmm4, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm4, [esp + nb134_qqMH] ;# vcoul mulsd xmm0, xmm4 addsd xmm6, xmm4 movsd xmm1, xmm0 movsd xmm2, xmm0 movapd xmm3, [esp + nb134_fjxH1] movapd xmm4, [esp + nb134_fjyH1] movapd xmm5, [esp + nb134_fjzH1] mulsd xmm0, [esp + nb134_dxMH1] mulsd xmm1, [esp + nb134_dyMH1] mulsd xmm2, [esp + nb134_dzMH1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb134_fixM] addsd xmm1, [esp + nb134_fiyM] addsd xmm2, [esp + nb134_fizM] movsd [esp + nb134_fjxH1], xmm3 movsd [esp + nb134_fjyH1], xmm4 movsd [esp + nb134_fjzH1], xmm5 movsd [esp + nb134_fixM], xmm0 movsd [esp + nb134_fiyM], xmm1 movsd [esp + nb134_fizM], xmm2 ;# M-H2 interaction movsd xmm0, [esp + nb134_rinvMH2] movsd xmm4, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm4, [esp + nb134_qqMH] ;# vcoul mulsd xmm0, xmm4 addsd xmm6, xmm4 movsd xmm1, xmm0 movsd xmm2, xmm0 movapd xmm3, [esp + nb134_fjxH2] movapd xmm4, [esp + nb134_fjyH2] movapd xmm5, [esp + nb134_fjzH2] mulsd xmm0, [esp + nb134_dxMH2] mulsd xmm1, [esp + nb134_dyMH2] mulsd xmm2, [esp + nb134_dzMH2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb134_fixM] addsd xmm1, [esp + nb134_fiyM] addsd xmm2, [esp + nb134_fizM] movsd [esp + nb134_fjxH2], xmm3 movsd [esp + nb134_fjyH2], xmm4 movsd [esp + nb134_fjzH2], xmm5 movsd [esp + nb134_fixM], xmm0 movsd [esp + nb134_fiyM], xmm1 movsd [esp + nb134_fizM], xmm2 ;# M-M interaction movsd xmm0, [esp + nb134_rinvMM] movsd xmm4, xmm0 ;# xmm7=rinv mulsd xmm0, xmm0 ;# xmm0=rinvsq mulsd xmm4, [esp + nb134_qqMM] ;# vcoul mulsd xmm0, xmm4 addsd xmm6, xmm4 movsd [esp + nb134_vctot], xmm6 movsd xmm1, xmm0 movsd xmm2, xmm0 movapd xmm3, [esp + nb134_fjxM] movapd xmm4, [esp + nb134_fjyM] movapd xmm5, [esp + nb134_fjzM] mulsd xmm0, [esp + nb134_dxMM] mulsd xmm1, [esp + nb134_dyMM] mulsd xmm2, [esp + nb134_dzMM] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb134_fixM] addsd xmm1, [esp + nb134_fiyM] addsd xmm2, [esp + nb134_fizM] movsd [esp + nb134_fjxM], xmm3 movsd [esp + nb134_fjyM], xmm4 movsd [esp + nb134_fjzM], xmm5 movsd [esp + nb134_fixM], xmm0 movsd [esp + nb134_fiyM], xmm1 movsd [esp + nb134_fizM], xmm2 mov edi, [ebp + nb134_faction] ;# Did all interactions - now update j forces ;# Step1 - merge forces movlpd xmm0, [esp + nb134_fjxO] movlpd xmm1, [esp + nb134_fjzO] movlpd xmm2, [esp + nb134_fjyH1] movlpd xmm3, [esp + nb134_fjxH2] movlpd xmm4, [esp + nb134_fjzH2] movlpd xmm5, [esp + nb134_fjyM] movhpd xmm0, [esp + nb134_fjyO] movhpd xmm1, [esp + nb134_fjxH1] movhpd xmm2, [esp + nb134_fjzH1] movhpd xmm3, [esp + nb134_fjyH2] movhpd xmm4, [esp + nb134_fjxM] movhpd xmm5, [esp + nb134_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 movlpd [edi + eax*8], xmm0 movhpd [edi + eax*8 + 8], xmm0 movlpd [edi + eax*8 + 16], xmm1 movhpd [edi + eax*8 + 24], xmm1 movlpd [edi + eax*8 + 32], xmm2 movhpd [edi + eax*8 + 40], xmm2 movlpd [edi + eax*8 + 48], xmm3 movhpd [edi + eax*8 + 56], xmm3 movlpd [edi + eax*8 + 64], xmm4 movhpd [edi + eax*8 + 72], xmm4 movlpd [edi + eax*8 + 80], xmm5 movhpd [edi + eax*8 + 88], xmm5 .nb134_updateouterdata: mov ecx, [esp + nb134_ii3] mov edi, [ebp + nb134_faction] mov esi, [ebp + nb134_fshift] mov edx, [esp + nb134_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movapd xmm0, [esp + nb134_fixO] movapd xmm1, [esp + nb134_fiyO] movapd xmm2, [esp + nb134_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
⌨️ 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?