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
mulpd xmm1, [esp + nb204_dyMH2] mulpd xmm2, [esp + nb204_dzMH2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb204_fixM] addpd xmm1, [esp + nb204_fiyM] addpd xmm2, [esp + nb204_fizM] movapd [esp + nb204_fjxH2], xmm3 movapd [esp + nb204_fjyH2], xmm4 movapd [esp + nb204_fjzH2], xmm5 movapd [esp + nb204_fixM], xmm0 movapd [esp + nb204_fiyM], xmm1 movapd [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 mulpd xmm5, [esp + nb204_rsqH1M] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=rinv+ krsq mulpd xmm0, xmm0 subpd xmm4, [esp + nb204_crf] mulpd xmm4, [esp + nb204_qqMH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb204_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb204_qqMH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxM] movapd xmm4, [esp + nb204_fjyM] movapd xmm5, [esp + nb204_fjzM] mulpd xmm0, [esp + nb204_dxH1M] mulpd xmm1, [esp + nb204_dyH1M] mulpd xmm2, [esp + nb204_dzH1M] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb204_fixH1] addpd xmm1, [esp + nb204_fiyH1] addpd xmm2, [esp + nb204_fizH1] movapd [esp + nb204_fjxM], xmm3 movapd [esp + nb204_fjyM], xmm4 movapd [esp + nb204_fjzM], xmm5 movapd [esp + nb204_fixH1], xmm0 movapd [esp + nb204_fiyH1], xmm1 movapd [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 mulpd xmm5, [esp + nb204_rsqH1H1] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb204_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb204_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb204_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb204_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxH1] movapd xmm4, [esp + nb204_fjyH1] movapd xmm5, [esp + nb204_fjzH1] mulpd xmm0, [esp + nb204_dxH1H1] mulpd xmm1, [esp + nb204_dyH1H1] mulpd xmm2, [esp + nb204_dzH1H1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb204_fixH1] addpd xmm1, [esp + nb204_fiyH1] addpd xmm2, [esp + nb204_fizH1] movapd [esp + nb204_fjxH1], xmm3 movapd [esp + nb204_fjyH1], xmm4 movapd [esp + nb204_fjzH1], xmm5 movapd [esp + nb204_fixH1], xmm0 movapd [esp + nb204_fiyH1], xmm1 movapd [esp + nb204_fizH1], xmm2 ;# H1-H2 interaction movapd xmm0, [esp + nb204_rinvH1H2] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb204_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb204_rsqH1H2] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq mulpd xmm0, xmm0 subpd xmm4, [esp + nb204_crf] mulpd xmm4, [esp + nb204_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb204_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb204_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxH2] movapd xmm4, [esp + nb204_fjyH2] movapd xmm5, [esp + nb204_fjzH2] mulpd xmm0, [esp + nb204_dxH1H2] mulpd xmm1, [esp + nb204_dyH1H2] mulpd xmm2, [esp + nb204_dzH1H2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb204_fixH1] addpd xmm1, [esp + nb204_fiyH1] addpd xmm2, [esp + nb204_fizH1] movapd [esp + nb204_fjxH2], xmm3 movapd [esp + nb204_fjyH2], xmm4 movapd [esp + nb204_fjzH2], xmm5 movapd [esp + nb204_fixH1], xmm0 movapd [esp + nb204_fiyH1], xmm1 movapd [esp + nb204_fizH1], xmm2 ;# H2-M interaction movapd xmm0, [esp + nb204_rinvH2M] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb204_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb204_rsqH2M] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb204_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb204_qqMH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb204_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb204_qqMH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxM] movapd xmm4, [esp + nb204_fjyM] movapd xmm5, [esp + nb204_fjzM] mulpd xmm0, [esp + nb204_dxH2M] mulpd xmm1, [esp + nb204_dyH2M] mulpd xmm2, [esp + nb204_dzH2M] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb204_fixH2] addpd xmm1, [esp + nb204_fiyH2] addpd xmm2, [esp + nb204_fizH2] movapd [esp + nb204_fjxM], xmm3 movapd [esp + nb204_fjyM], xmm4 movapd [esp + nb204_fjzM], xmm5 movapd [esp + nb204_fixH2], xmm0 movapd [esp + nb204_fiyH2], xmm1 movapd [esp + nb204_fizH2], xmm2 ;# H2-H1 interaction movapd xmm0, [esp + nb204_rinvH2H1] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb204_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb204_rsqH2H1] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb204_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb204_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb204_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb204_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxH1] movapd xmm4, [esp + nb204_fjyH1] movapd xmm5, [esp + nb204_fjzH1] mulpd xmm0, [esp + nb204_dxH2H1] mulpd xmm1, [esp + nb204_dyH2H1] mulpd xmm2, [esp + nb204_dzH2H1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb204_fixH2] addpd xmm1, [esp + nb204_fiyH2] addpd xmm2, [esp + nb204_fizH2] movapd [esp + nb204_fjxH1], xmm3 movapd [esp + nb204_fjyH1], xmm4 movapd [esp + nb204_fjzH1], xmm5 movapd [esp + nb204_fixH2], xmm0 movapd [esp + nb204_fiyH2], xmm1 movapd [esp + nb204_fizH2], xmm2 ;# H2-H2 interaction movapd xmm0, [esp + nb204_rinvH2H2] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb204_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb204_rsqH2H2] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb204_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb204_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb204_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb204_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm1, xmm0 movapd [esp + nb204_vctot], xmm6 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxH2] movapd xmm4, [esp + nb204_fjyH2] movapd xmm5, [esp + nb204_fjzH2] mulpd xmm0, [esp + nb204_dxH2H2] mulpd xmm1, [esp + nb204_dyH2H2] mulpd xmm2, [esp + nb204_dzH2H2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb204_fixH2] addpd xmm1, [esp + nb204_fiyH2] addpd xmm2, [esp + nb204_fizH2] movapd [esp + nb204_fjxH2], xmm3 movapd [esp + nb204_fjyH2], xmm4 movapd [esp + nb204_fjzH2], xmm5 movapd [esp + nb204_fixH2], xmm0 movapd [esp + nb204_fiyH2], xmm1 movapd [esp + nb204_fizH2], xmm2 mov edi, [ebp + nb204_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] movhpd xmm0, [edi + ebx*8 + 24] movhpd xmm1, [edi + ebx*8 + 32] movhpd xmm2, [edi + ebx*8 + 40] movhpd xmm3, [edi + ebx*8 + 48] movhpd xmm4, [edi + ebx*8 + 56] movhpd xmm5, [edi + ebx*8 + 64] movhpd xmm6, [edi + ebx*8 + 72] movhpd xmm7, [edi + ebx*8 + 80] addpd xmm0, [esp + nb204_fjxH1] addpd xmm1, [esp + nb204_fjyH1] addpd xmm2, [esp + nb204_fjzH1] addpd xmm3, [esp + nb204_fjxH2] addpd xmm4, [esp + nb204_fjyH2] addpd xmm5, [esp + nb204_fjzH2] addpd xmm6, [esp + nb204_fjxM] addpd xmm7, [esp + nb204_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 movhpd [edi + ebx*8 + 24], xmm0 movhpd [edi + ebx*8 + 32], xmm1 movhpd [edi + ebx*8 + 40], xmm2 movhpd [edi + ebx*8 + 48], xmm3 movhpd [edi + ebx*8 + 56], xmm4 movhpd [edi + ebx*8 + 64], xmm5 movhpd [edi + ebx*8 + 72], xmm6 movhpd [edi + ebx*8 + 80], xmm7 movlpd xmm0, [edi + eax*8 + 88] movhpd xmm0, [edi + ebx*8 + 88] addpd xmm0, [esp + nb204_fjzM] movlpd [edi + eax*8 + 88], xmm0 movhpd [edi + ebx*8 + 88], xmm0 ;# should we do one more iteration? sub dword ptr [esp + nb204_innerk], 2 jl .nb204_checksingle jmp .nb204_unroll_loop.nb204_checksingle: mov edx, [esp + nb204_innerk] and edx, 1 jnz .nb204_dosingle jmp .nb204_updateouterdata.nb204_dosingle: mov edx, [esp + nb204_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] mov esi, [ebp + nb204_pos] lea eax, [eax + eax*2] ;# move j coordinates to local temp variables movlpd xmm2, [esi + eax*8 + 24] movlpd xmm3, [esi + eax*8 + 32] movlpd xmm4, [esi + eax*8 + 40] movlpd xmm5, [esi + eax*8 + 48] movlpd xmm6, [esi + eax*8 + 56] movlpd xmm7, [esi + eax*8 + 64] movapd [esp + nb204_jxH1], xmm2 movapd [esp + nb204_jyH1], xmm3 movapd [esp + nb204_jzH1], xmm4 movapd [esp + nb204_jxH2], xmm5 movapd [esp + nb204_jyH2], xmm6 movapd [esp + nb204_jzH2], xmm7 movlpd xmm2, [esi + eax*8 + 72] movlpd xmm3, [esi + eax*8 + 80] movlpd xmm4, [esi + eax*8 + 88] movapd [esp + nb204_jxM], xmm2 movapd [esp + nb204_jyM], xmm3 movapd [esp + nb204_jzM], xmm4 movapd xmm0, [esp + nb204_ixM] movapd xmm1, [esp + nb204_iyM] movapd xmm2, [esp + nb204_izM] movapd xmm3, [esp + nb204_ixM] movapd xmm4, [esp + nb204_iyM] movapd xmm5, [esp + nb204_izM] 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_dxMM], xmm0 movapd [esp + nb204_dyMM], xmm1 movapd [esp + nb204_dzMM], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb204_dxMH1], xmm3 movapd [esp + nb204_dyMH1], xmm4 movapd [esp + nb204_dzMH1], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd xmm3, xmm5 movapd [esp + nb204_rsqMM], xmm0 movapd [esp + nb204_rsqMH1], xmm3 movapd xmm0, [esp + nb204_ixM] movapd xmm1, [esp + nb204_iyM] movapd xmm2, [esp + nb204_izM] movapd xmm3, [esp + nb204_ixH1] movapd xmm4, [esp + nb204_iyH1] movapd xmm5, [esp + nb204_izH1] subsd xmm0, [esp + nb204_jxH2] subsd xmm1, [esp + nb204_jyH2] subsd xmm2, [esp + nb204_jzH2] subsd xmm3, [esp + nb204_jxM] subsd xmm4, [esp + nb204_jyM] subsd xmm5, [esp + nb204_jzM] movapd [esp + nb204_dxMH2], xmm0 movapd [esp + nb204_dyMH2], xmm1 movapd [esp + nb204_dzMH2], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb204_dxH1M], xmm3 movapd [esp + nb204_dyH1M], xmm4 movapd [esp + nb204_dzH1M], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd xmm3, xmm5 movapd [esp + nb204_rsqMH2], xmm0 movapd [esp + nb204_rsqH1M], xmm3 movapd xmm0, [esp + nb204_ixH1] movapd xmm1, [esp + nb204_iyH1] movapd xmm2, [esp + nb204_izH1] movapd xmm3, [esp + nb204_ixH1] movapd xmm4, [esp + nb204_iyH1] movapd xmm5, [esp + nb204_izH1] subsd xmm0, [esp + nb204_jxH1] subsd xmm1, [esp + nb204_jyH1] subsd xmm2, [esp + nb204_jzH1] subsd xmm3, [esp + nb204_jxH2] subsd xmm4, [esp + nb204_jyH2] subsd xmm5, [esp + nb204_jzH2] movapd [esp + nb204_dxH1H1], xmm0 movapd [esp + nb204_dyH1H1], xmm1 movapd [esp + nb204_dzH1H1], xmm2 mulsd xmm0, xmm0 mulsd xmm1, xmm1 mulsd xmm2, xmm2 movapd [esp + nb204_dxH1H2], xmm3 movapd [esp + nb204_dyH1H2], xmm4 movapd [esp + nb204_dzH1H2], xmm5 mulsd xmm3, xmm3 mulsd xmm4, xmm4 mulsd xmm5, xmm5 addsd xmm0, xmm1 addsd xmm0, xmm2 addsd xmm3, xmm4 addsd xmm3, xmm5 movapd [esp + nb204_rsqH1H1], xmm0
⌨️ 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?