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
mulpd xmm1, [esp + nb114_dyOO] mulpd xmm2, [esp + nb114_dzOO] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb114_fixO] addpd xmm1, [esp + nb114_fiyO] addpd xmm2, [esp + nb114_fizO] movapd [esp + nb114_fjxO], xmm3 movapd [esp + nb114_fjyO], xmm4 movapd [esp + nb114_fjzO], xmm5 movapd [esp + nb114_fixO], xmm0 movapd [esp + nb114_fiyO], xmm1 movapd [esp + nb114_fizO], xmm2 ;# H1-H1 interaction movapd xmm0, [esp + nb114_rinvH1H1] movapd xmm7, xmm0 ;# xmm7=rinv mulpd xmm0, xmm0 ;# xmm0=rinvsq mulpd xmm7, [esp + nb114_qqHH] ;# xmm7=vcoul mulpd xmm0, xmm7 ;# fscal addpd xmm7, [esp + nb114_vctot] ;# local vctot summation variable movapd [esp + nb114_vctot], xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb114_dxH1H1] mulpd xmm1, [esp + nb114_dyH1H1] mulpd xmm2, [esp + nb114_dzH1H1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb114_fixH1] addpd xmm1, [esp + nb114_fiyH1] addpd xmm2, [esp + nb114_fizH1] movapd [esp + nb114_fjxH1], xmm3 movapd [esp + nb114_fjyH1], xmm4 movapd [esp + nb114_fjzH1], xmm5 movapd [esp + nb114_fixH1], xmm0 movapd [esp + nb114_fiyH1], xmm1 movapd [esp + nb114_fizH1], xmm2 ;# H1-H2 interaction movapd xmm0, [esp + nb114_rinvH1H2] movapd xmm7, xmm0 ;# xmm7=rinv mulpd xmm0, xmm0 ;# xmm0=rinvsq mulpd xmm7, [esp + nb114_qqHH] ;# vcoul mulpd xmm0, xmm7 addpd xmm7, [esp + nb114_vctot] ;# local vctot summation variable movapd [esp + nb114_vctot], xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb114_dxH1H2] mulpd xmm1, [esp + nb114_dyH1H2] mulpd xmm2, [esp + nb114_dzH1H2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb114_fixH1] addpd xmm1, [esp + nb114_fiyH1] addpd xmm2, [esp + nb114_fizH1] movapd [esp + nb114_fjxH2], xmm3 movapd [esp + nb114_fjyH2], xmm4 movapd [esp + nb114_fjzH2], xmm5 movapd [esp + nb114_fixH1], xmm0 movapd [esp + nb114_fiyH1], xmm1 movapd [esp + nb114_fizH1], xmm2 ;# H1-M interaction movapd xmm0, [esp + nb114_rinvH1M] movapd xmm7, xmm0 ;# xmm7=rinv mulpd xmm0, xmm0 ;# xmm0=rinvsq mulpd xmm7, [esp + nb114_qqMH] ;# vcoul mulpd xmm0, xmm7 addpd xmm7, [esp + nb114_vctot] ;# local vctot summation variable movapd [esp + nb114_vctot], xmm7 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb114_dxH1M] mulpd xmm1, [esp + nb114_dyH1M] mulpd xmm2, [esp + nb114_dzH1M] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb114_fixH1] addpd xmm1, [esp + nb114_fiyH1] addpd xmm2, [esp + nb114_fizH1] movapd [esp + nb114_fjxM], xmm3 movapd [esp + nb114_fjyM], xmm4 movapd [esp + nb114_fjzM], xmm5 movapd [esp + nb114_fixH1], xmm0 movapd [esp + nb114_fiyH1], xmm1 movapd [esp + nb114_fizH1], xmm2 ;# H2-H1 interaction movapd xmm0, [esp + nb114_rinvH2H1] movapd xmm7, xmm0 ;# xmm7=rinv mulpd xmm0, xmm0 ;# xmm0=rinvsq mulpd xmm7, [esp + nb114_qqHH] mulpd xmm0, xmm7 addpd xmm7, [esp + nb114_vctot] ;# local vctot summation variable movapd [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] mulpd xmm0, [esp + nb114_dxH2H1] mulpd xmm1, [esp + nb114_dyH2H1] mulpd xmm2, [esp + nb114_dzH2H1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb114_fixH2] addpd xmm1, [esp + nb114_fiyH2] addpd xmm2, [esp + nb114_fizH2] movapd [esp + nb114_fjxH1], xmm3 movapd [esp + nb114_fjyH1], xmm4 movapd [esp + nb114_fjzH1], xmm5 movapd [esp + nb114_fixH2], xmm0 movapd [esp + nb114_fiyH2], xmm1 movapd [esp + nb114_fizH2], xmm2 ;# H2-H2 interaction movapd xmm0, [esp + nb114_rinvH2H2] movapd xmm7, xmm0 ;# xmm7=rinv mulpd xmm0, xmm0 ;# xmm0=rinvsq mulpd xmm7, [esp + nb114_qqHH] ;# vcoul mulpd xmm0, xmm7 addpd xmm7, [esp + nb114_vctot] ;# local vctot summation variable movapd [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] mulpd xmm0, [esp + nb114_dxH2H2] mulpd xmm1, [esp + nb114_dyH2H2] mulpd xmm2, [esp + nb114_dzH2H2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb114_fixH2] addpd xmm1, [esp + nb114_fiyH2] addpd xmm2, [esp + nb114_fizH2] movapd [esp + nb114_fjxH2], xmm3 movapd [esp + nb114_fjyH2], xmm4 movapd [esp + nb114_fjzH2], xmm5 movapd [esp + nb114_fixH2], xmm0 movapd [esp + nb114_fiyH2], xmm1 movapd [esp + nb114_fizH2], xmm2 ;# H2-M interaction movapd xmm0, [esp + nb114_rinvH2M] movapd xmm7, xmm0 ;# xmm7=rinv mulpd xmm0, xmm0 ;# xmm0=rinvsq mulpd xmm7, [esp + nb114_qqMH] ;# vcoul mulpd xmm0, xmm7 addpd xmm7, [esp + nb114_vctot] movapd [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] mulpd xmm0, [esp + nb114_dxH2M] mulpd xmm1, [esp + nb114_dyH2M] mulpd xmm2, [esp + nb114_dzH2M] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb114_fixH2] addpd xmm1, [esp + nb114_fiyH2] addpd xmm2, [esp + nb114_fizH2] movapd [esp + nb114_fjxM], xmm3 movapd [esp + nb114_fjyM], xmm4 movapd [esp + nb114_fjzM], xmm5 movapd [esp + nb114_fixH2], xmm0 movapd [esp + nb114_fiyH2], xmm1 movapd [esp + nb114_fizH2], xmm2 ;# M-H1 interaction movapd xmm0, [esp + nb114_rinvMH1] movapd xmm7, xmm0 ;# xmm7=rinv mulpd xmm0, xmm0 ;# xmm0=rinvsq mulpd xmm7, [esp + nb114_qqMH] mulpd xmm0, xmm7 addpd xmm7, [esp + nb114_vctot] ;# local vctot summation variable movapd [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] mulpd xmm0, [esp + nb114_dxMH1] mulpd xmm1, [esp + nb114_dyMH1] mulpd xmm2, [esp + nb114_dzMH1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb114_fixM] addpd xmm1, [esp + nb114_fiyM] addpd xmm2, [esp + nb114_fizM] movapd [esp + nb114_fjxH1], xmm3 movapd [esp + nb114_fjyH1], xmm4 movapd [esp + nb114_fjzH1], xmm5 movapd [esp + nb114_fixM], xmm0 movapd [esp + nb114_fiyM], xmm1 movapd [esp + nb114_fizM], xmm2 ;# M-H2 interaction movapd xmm0, [esp + nb114_rinvMH2] movapd xmm7, xmm0 ;# xmm7=rinv mulpd xmm0, xmm0 ;# xmm0=rinvsq mulpd xmm7, [esp + nb114_qqMH] mulpd xmm0, xmm7 addpd xmm7, [esp + nb114_vctot] movapd [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] mulpd xmm0, [esp + nb114_dxMH2] mulpd xmm1, [esp + nb114_dyMH2] mulpd xmm2, [esp + nb114_dzMH2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb114_fixM] addpd xmm1, [esp + nb114_fiyM] addpd xmm2, [esp + nb114_fizM] movapd [esp + nb114_fjxH2], xmm3 movapd [esp + nb114_fjyH2], xmm4 movapd [esp + nb114_fjzH2], xmm5 movapd [esp + nb114_fixM], xmm0 movapd [esp + nb114_fiyM], xmm1 movapd [esp + nb114_fizM], xmm2 ;# M-M interaction movapd xmm0, [esp + nb114_rinvMM] movapd xmm7, xmm0 ;# xmm7=rinv mulpd xmm0, xmm0 ;# xmm0=rinvsq mulpd xmm7, [esp + nb114_qqMM] ;# vcoul mulpd xmm0, xmm7 addpd xmm7, [esp + nb114_vctot] ;# local vctot summation variable movapd [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] mulpd xmm0, [esp + nb114_dxMM] mulpd xmm1, [esp + nb114_dyMM] mulpd xmm2, [esp + nb114_dzMM] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb114_fixM] addpd xmm1, [esp + nb114_fiyM] addpd xmm2, [esp + nb114_fizM] movapd [esp + nb114_fjxM], xmm3 movapd [esp + nb114_fjyM], xmm4 movapd [esp + nb114_fjzM], xmm5 movapd [esp + nb114_fixM], xmm0 movapd [esp + nb114_fiyM], xmm1 movapd [esp + nb114_fizM], xmm2 mov edi, [ebp + nb114_faction] ;# Did all interactions - now update j forces ;# Step1 - transpose fjxO, fjyO and fjzO, fjxH1 movapd xmm0, [esp + nb114_fjxO] movapd xmm1, [esp + nb114_fjyO] movapd xmm2, [esp + nb114_fjzO] movapd xmm3, [esp + nb114_fjxH1] movapd xmm4, xmm0 movapd xmm5, xmm2 unpcklpd xmm0, xmm1 ;# fjOxA fjOyA unpckhpd xmm4, xmm1 ;# fjOxB fjOyB unpcklpd xmm2, xmm3 ;# fjOzA fjH1xA unpckhpd xmm5, xmm3 ;# fjOzB fjH1xB movlpd xmm1, [edi + eax*8] movlpd xmm3, [edi + ebx*8] movhpd xmm1, [edi + eax*8 + 8] movhpd xmm3, [edi + ebx*8 + 8] movlpd xmm6, [edi + eax*8 + 16] movlpd xmm7, [edi + ebx*8 + 16] movhpd xmm6, [edi + eax*8 + 24] movhpd xmm7, [edi + ebx*8 + 24] addpd xmm1, xmm0 addpd xmm3, xmm4 addpd xmm6, xmm2 addpd xmm7, xmm5 movlpd [edi + eax*8], xmm1 movlpd [edi + ebx*8], xmm3 movhpd [edi + eax*8 + 8], xmm1 movhpd [edi + ebx*8 + 8], xmm3 movlpd [edi + eax*8 + 16], xmm6 movlpd [edi + ebx*8 + 16], xmm7 movhpd [edi + eax*8 + 24], xmm6 movhpd [edi + ebx*8 + 24], xmm7 ;# Step2 - transpose fjyH1, fjzH1 and fjxH2, fjyH2 movapd xmm0, [esp + nb114_fjyH1] movapd xmm1, [esp + nb114_fjzH1] movapd xmm2, [esp + nb114_fjxH2] movapd xmm3, [esp + nb114_fjyH2] movapd xmm4, xmm0 movapd xmm5, xmm2 unpcklpd xmm0, xmm1 ;# fjOxA fjOyA unpckhpd xmm4, xmm1 ;# fjOxB fjOyB unpcklpd xmm2, xmm3 ;# fjOzA fjH1xA unpckhpd xmm5, xmm3 ;# fjOzB fjH1xB movlpd xmm1, [edi + eax*8 + 32] movlpd xmm3, [edi + ebx*8 + 32] movhpd xmm1, [edi + eax*8 + 40] movhpd xmm3, [edi + ebx*8 + 40] movlpd xmm6, [edi + eax*8 + 48] movlpd xmm7, [edi + ebx*8 + 48] movhpd xmm6, [edi + eax*8 + 56] movhpd xmm7, [edi + ebx*8 + 56] addpd xmm1, xmm0 addpd xmm3, xmm4 addpd xmm6, xmm2 addpd xmm7, xmm5 movlpd [edi + eax*8 + 32], xmm1 movlpd [edi + ebx*8 + 32], xmm3 movhpd [edi + eax*8 + 40], xmm1 movhpd [edi + ebx*8 + 40], xmm3 movlpd [edi + eax*8 + 48], xmm6 movlpd [edi + ebx*8 + 48], xmm7 movhpd [edi + eax*8 + 56], xmm6 movhpd [edi + ebx*8 + 56], xmm7 ;# Step3 - transpose fjzH2, fjxM and fjyM, fjzM movapd xmm0, [esp + nb114_fjzH2] movapd xmm1, [esp + nb114_fjxM] movapd xmm2, [esp + nb114_fjyM] movapd xmm3, [esp + nb114_fjzM] movapd xmm4, xmm0 movapd xmm5, xmm2 unpcklpd xmm0, xmm1 ;# fjOxA fjOyA unpckhpd xmm4, xmm1 ;# fjOxB fjOyB unpcklpd xmm2, xmm3 ;# fjOzA fjH1xA unpckhpd xmm5, xmm3 ;# fjOzB fjH1xB movlpd xmm1, [edi + eax*8 + 64] movlpd xmm3, [edi + ebx*8 + 64] movhpd xmm1, [edi + eax*8 + 72] movhpd xmm3, [edi + ebx*8 + 72] movlpd xmm6, [edi + eax*8 + 80] movlpd xmm7, [edi + ebx*8 + 80] movhpd xmm6, [edi + eax*8 + 88] movhpd xmm7, [edi + ebx*8 + 88] addpd xmm1, xmm0 addpd xmm3, xmm4 addpd xmm6, xmm2 addpd xmm7, xmm5 movlpd [edi + eax*8 + 64], xmm1 movlpd [edi + ebx*8 + 64], xmm3 movhpd [edi + eax*8 + 72], xmm1 movhpd [edi + ebx*8 + 72], xmm3 movlpd [edi + eax*8 + 80], xmm6 movlpd [edi + ebx*8 + 80], xmm7 movhpd [edi + eax*8 + 88], xmm6 movhpd [edi + ebx*8 + 88], xmm7 ;# should we do one more iteration? sub dword ptr [esp + nb114_innerk], 2 jl .nb114_checksingle jmp .nb114_unroll_loop.nb114_checksingle: mov edx, [esp + nb114_innerk] and edx, 1 jnz .nb114_dosingle jmp .nb114_updateouterdata.nb114_dosingle: mov edx, [esp + nb114_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] mov esi, [ebp + nb114_pos] ;# base of pos[] lea eax, [eax + eax*2] ;# replace jnr with j3 ;# move j coordinates to local temp variables ;# load ox, oy, oz, h1x movlpd xmm0, [esi + eax*8] movhpd xmm0, [esi + eax*8 + 8] movlpd xmm1, [esi + eax*8 + 16] movhpd xmm1, [esi + eax*8 + 24] movlpd xmm2, [esi + eax*8 + 32] movhpd xmm2, [esi + eax*8 + 40] movlpd xmm3, [esi + eax*8 + 48] movhpd xmm3, [esi + eax*8 + 56] movlpd xmm4, [esi + eax*8 + 64] movhpd xmm4, [esi + eax*8 + 72] movlpd xmm5, [esi + eax*8 + 80] movhpd xmm5, [esi + eax*8 + 88] movsd [esp + nb114_jxO], xmm0 movsd [esp + nb114_jzO], xmm1 movsd [esp + nb114_jyH1], xmm2 movsd [esp + nb114_jxH2], xmm3 movsd [esp + nb114_jzH2], xmm4 movsd [esp + nb114_jyM], xmm5 unpckhpd xmm0, xmm0 unpckhpd xmm1, xmm1 unpckhpd xmm2, xmm2 unpckhpd xmm3, xmm3 unpckhpd xmm4, xmm4 unpckhpd xmm5, xmm5 movsd [esp + nb114_jyO], xmm0 movsd [esp + nb114_jxH1], xmm1 movsd [esp + nb114_jzH1], xmm2 movsd [esp + nb114_jyH2], xmm3
⌨️ 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?