nb_kernel232_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
movlpd xmm6, [esi + eax*8 + 48] ;# G1 movlpd xmm3, [esi + ebx*8 + 48] ;# G2 movhpd xmm6, [esi + eax*8 + 56] ;# G1 H1 movhpd xmm3, [esi + ebx*8 + 56] ;# G2 H2 movapd xmm7, xmm6 unpcklpd xmm6, xmm3 ;# G1 G2 unpckhpd xmm7, xmm3 ;# H1 H2 ;# table ready, in xmm4-xmm7 mulpd xmm6, xmm1 ;# xmm6=Geps mulpd xmm7, xmm2 ;# xmm7=Heps2 addpd xmm5, xmm6 addpd xmm5, xmm7 ;# xmm5=Fp mulpd xmm7, [esp + nb232_two] ;# two*Heps2 addpd xmm7, xmm6 addpd xmm7, xmm5 ;# xmm7=FF mulpd xmm5, xmm1 ;# xmm5=eps*Fp addpd xmm5, xmm4 ;# xmm5=VV movapd xmm4, [esp + nb232_c12] mulpd xmm7, xmm4 mulpd xmm5, xmm4 addpd xmm5, [esp + nb232_Vvdwtot] movapd xmm3, [esp + nb232_fstmp] mulpd xmm7, [esp + nb232_tsc] subpd xmm3, xmm7 movapd [esp + nb232_Vvdwtot], xmm5 mulpd xmm0, xmm3 movapd xmm1, xmm0 movapd xmm2, xmm0 movd eax, mm0 movd ebx, mm1 mov edi, [ebp + nb232_faction] ;# update forces xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb232_dxOO] mulpd xmm1, [esp + nb232_dyOO] mulpd xmm2, [esp + nb232_dzOO] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb232_fixO] addpd xmm1, [esp + nb232_fiyO] addpd xmm2, [esp + nb232_fizO] movapd [esp + nb232_fjxO], xmm3 movapd [esp + nb232_fjyO], xmm4 movapd [esp + nb232_fjzO], xmm5 movapd [esp + nb232_fixO], xmm0 movapd [esp + nb232_fiyO], xmm1 movapd [esp + nb232_fizO], xmm2 ;# O-H1 interaction movapd xmm0, [esp + nb232_rinvOH1] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb232_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb232_rsqOH1] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=rinv+ krsq mulpd xmm0, xmm0 subpd xmm4, [esp + nb232_crf] mulpd xmm4, [esp + nb232_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb232_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb232_qqOH] ;# xmm7 = coul part of fscal movapd xmm6, [esp + nb232_vctot] addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH1 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb232_dxOH1] mulpd xmm1, [esp + nb232_dyOH1] mulpd xmm2, [esp + nb232_dzOH1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb232_fixO] addpd xmm1, [esp + nb232_fiyO] addpd xmm2, [esp + nb232_fizO] movapd [esp + nb232_fjxH1], xmm3 movapd [esp + nb232_fjyH1], xmm4 movapd [esp + nb232_fjzH1], xmm5 movapd [esp + nb232_fixO], xmm0 movapd [esp + nb232_fiyO], xmm1 movapd [esp + nb232_fizO], xmm2 ;# O-H2 interaction movapd xmm0, [esp + nb232_rinvOH2] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb232_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb232_rsqOH2] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq mulpd xmm0, xmm0 subpd xmm4, [esp + nb232_crf] mulpd xmm4, [esp + nb232_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb232_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb232_qqOH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 xorpd xmm3, xmm3 movapd xmm4, xmm3 movapd xmm5, xmm3 mulpd xmm0, [esp + nb232_dxOH2] mulpd xmm1, [esp + nb232_dyOH2] mulpd xmm2, [esp + nb232_dzOH2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb232_fixO] addpd xmm1, [esp + nb232_fiyO] addpd xmm2, [esp + nb232_fizO] movapd [esp + nb232_fjxH2], xmm3 movapd [esp + nb232_fjyH2], xmm4 movapd [esp + nb232_fjzH2], xmm5 movapd [esp + nb232_fixO], xmm0 movapd [esp + nb232_fiyO], xmm1 movapd [esp + nb232_fizO], xmm2 ;# H1-O interaction movapd xmm0, [esp + nb232_rinvH1O] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb232_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb232_rsqH1O] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=rinv+ krsq mulpd xmm0, xmm0 subpd xmm4, [esp + nb232_crf] mulpd xmm4, [esp + nb232_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb232_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb232_qqOH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb232_fjxO] movapd xmm4, [esp + nb232_fjyO] movapd xmm5, [esp + nb232_fjzO] mulpd xmm0, [esp + nb232_dxH1O] mulpd xmm1, [esp + nb232_dyH1O] mulpd xmm2, [esp + nb232_dzH1O] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb232_fixH1] addpd xmm1, [esp + nb232_fiyH1] addpd xmm2, [esp + nb232_fizH1] movapd [esp + nb232_fjxO], xmm3 movapd [esp + nb232_fjyO], xmm4 movapd [esp + nb232_fjzO], xmm5 movapd [esp + nb232_fixH1], xmm0 movapd [esp + nb232_fiyH1], xmm1 movapd [esp + nb232_fizH1], xmm2 ;# H1-H1 interaction movapd xmm0, [esp + nb232_rinvH1H1] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb232_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb232_rsqH1H1] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb232_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb232_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb232_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb232_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb232_fjxH1] movapd xmm4, [esp + nb232_fjyH1] movapd xmm5, [esp + nb232_fjzH1] mulpd xmm0, [esp + nb232_dxH1H1] mulpd xmm1, [esp + nb232_dyH1H1] mulpd xmm2, [esp + nb232_dzH1H1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb232_fixH1] addpd xmm1, [esp + nb232_fiyH1] addpd xmm2, [esp + nb232_fizH1] movapd [esp + nb232_fjxH1], xmm3 movapd [esp + nb232_fjyH1], xmm4 movapd [esp + nb232_fjzH1], xmm5 movapd [esp + nb232_fixH1], xmm0 movapd [esp + nb232_fiyH1], xmm1 movapd [esp + nb232_fizH1], xmm2 ;# H1-H2 interaction movapd xmm0, [esp + nb232_rinvH1H2] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb232_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb232_rsqH1H2] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq mulpd xmm0, xmm0 subpd xmm4, [esp + nb232_crf] mulpd xmm4, [esp + nb232_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb232_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb232_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb232_fjxH2] movapd xmm4, [esp + nb232_fjyH2] movapd xmm5, [esp + nb232_fjzH2] mulpd xmm0, [esp + nb232_dxH1H2] mulpd xmm1, [esp + nb232_dyH1H2] mulpd xmm2, [esp + nb232_dzH1H2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb232_fixH1] addpd xmm1, [esp + nb232_fiyH1] addpd xmm2, [esp + nb232_fizH1] movapd [esp + nb232_fjxH2], xmm3 movapd [esp + nb232_fjyH2], xmm4 movapd [esp + nb232_fjzH2], xmm5 movapd [esp + nb232_fixH1], xmm0 movapd [esp + nb232_fiyH1], xmm1 movapd [esp + nb232_fizH1], xmm2 ;# H2-O interaction movapd xmm0, [esp + nb232_rinvH2O] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb232_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb232_rsqH2O] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb232_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb232_qqOH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb232_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb232_qqOH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb232_fjxO] movapd xmm4, [esp + nb232_fjyO] movapd xmm5, [esp + nb232_fjzO] mulpd xmm0, [esp + nb232_dxH2O] mulpd xmm1, [esp + nb232_dyH2O] mulpd xmm2, [esp + nb232_dzH2O] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb232_fixH2] addpd xmm1, [esp + nb232_fiyH2] addpd xmm2, [esp + nb232_fizH2] movapd [esp + nb232_fjxO], xmm3 movapd [esp + nb232_fjyO], xmm4 movapd [esp + nb232_fjzO], xmm5 movapd [esp + nb232_fixH2], xmm0 movapd [esp + nb232_fiyH2], xmm1 movapd [esp + nb232_fizH2], xmm2 ;# H2-H1 interaction movapd xmm0, [esp + nb232_rinvH2H1] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb232_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb232_rsqH2H1] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb232_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb232_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb232_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb232_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb232_fjxH1] movapd xmm4, [esp + nb232_fjyH1] movapd xmm5, [esp + nb232_fjzH1] mulpd xmm0, [esp + nb232_dxH2H1] mulpd xmm1, [esp + nb232_dyH2H1] mulpd xmm2, [esp + nb232_dzH2H1] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb232_fixH2] addpd xmm1, [esp + nb232_fiyH2] addpd xmm2, [esp + nb232_fizH2] movapd [esp + nb232_fjxH1], xmm3 movapd [esp + nb232_fjyH1], xmm4 movapd [esp + nb232_fjzH1], xmm5 movapd [esp + nb232_fixH2], xmm0 movapd [esp + nb232_fiyH2], xmm1 movapd [esp + nb232_fizH2], xmm2 ;# H2-H2 interaction movapd xmm0, [esp + nb232_rinvH2H2] movapd xmm7, xmm0 ;# xmm7=rinv movapd xmm5, [esp + nb232_krf] movapd xmm1, xmm0 mulpd xmm5, [esp + nb232_rsqH2H2] ;# xmm5=krsq movapd xmm4, xmm5 addpd xmm4, xmm7 ;# xmm4=r inv+ krsq subpd xmm4, [esp + nb232_crf] mulpd xmm0, xmm0 mulpd xmm4, [esp + nb232_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulpd xmm5, [esp + nb232_two] subpd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulpd xmm7, [esp + nb232_qqHH] ;# xmm7 = coul part of fscal addpd xmm6, xmm4 ;# add to local vctot mulpd xmm0, xmm7 ;# fsOH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm1, xmm0 movapd [esp + nb232_vctot], xmm6 movapd xmm2, xmm0 movapd xmm3, [esp + nb232_fjxH2] movapd xmm4, [esp + nb232_fjyH2] movapd xmm5, [esp + nb232_fjzH2] mulpd xmm0, [esp + nb232_dxH2H2] mulpd xmm1, [esp + nb232_dyH2H2] mulpd xmm2, [esp + nb232_dzH2H2] subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 addpd xmm0, [esp + nb232_fixH2] addpd xmm1, [esp + nb232_fiyH2] addpd xmm2, [esp + nb232_fizH2] movapd [esp + nb232_fjxH2], xmm3 movapd [esp + nb232_fjyH2], xmm4 movapd [esp + nb232_fjzH2], xmm5 movapd [esp + nb232_fixH2], xmm0 movapd [esp + nb232_fiyH2], xmm1 movapd [esp + nb232_fizH2], xmm2 mov edi, [ebp + nb232_faction] ;# Did all interactions - now update j forces movlpd xmm0, [edi + eax*8] movlpd xmm1, [edi + eax*8 + 8] movlpd xmm2, [edi + eax*8 + 16] movlpd xmm3, [edi + eax*8 + 24] movlpd xmm4, [edi + eax*8 + 32] movlpd xmm5, [edi + eax*8 + 40] movlpd xmm6, [edi + eax*8 + 48] movlpd xmm7, [edi + eax*8 + 56] movhpd xmm0, [edi + ebx*8] movhpd xmm1, [edi + ebx*8 + 8] movhpd xmm2, [edi + ebx*8 + 16] movhpd xmm3, [edi + ebx*8 + 24] movhpd xmm4, [edi + ebx*8 + 32] movhpd xmm5, [edi + ebx*8 + 40] movhpd xmm6, [edi + ebx*8 + 48] movhpd xmm7, [edi + ebx*8 + 56] addpd xmm0, [esp + nb232_fjxO] addpd xmm1, [esp + nb232_fjyO] addpd xmm2, [esp + nb232_fjzO] addpd xmm3, [esp + nb232_fjxH1] addpd xmm4, [esp + nb232_fjyH1] addpd xmm5, [esp + nb232_fjzH1] addpd xmm6, [esp + nb232_fjxH2] addpd xmm7, [esp + nb232_fjyH2] movlpd [edi + eax*8], xmm0 movlpd [edi + eax*8 + 8], xmm1 movlpd [edi + eax*8 + 16], xmm2 movlpd [edi + eax*8 + 24], xmm3 movlpd [edi + eax*8 + 32], xmm4 movlpd [edi + eax*8 + 40], xmm5 movlpd [edi + eax*8 + 48], xmm6 movlpd [edi + eax*8 + 56], xmm7 movhpd [edi + ebx*8], xmm0 movhpd [edi + ebx*8 + 8], xmm1 movhpd [edi + ebx*8 + 16], xmm2 movhpd [edi + ebx*8 + 24], xmm3 movhpd [edi + ebx*8 + 32], xmm4 movhpd [edi + ebx*8 + 40], xmm5 movhpd [edi + ebx*8 + 48], xmm6 movhpd [edi + ebx*8 + 56], xmm7 movlpd xmm0, [edi + eax*8 + 64] movhpd xmm0, [edi + ebx*8 + 64] addpd xmm0, [esp + nb232_fjzH2] movlpd [edi + eax*8 + 64], xmm0 movhpd [edi + ebx*8 + 64], xmm0 ;# should we do one more iteration? sub dword ptr [esp + nb232_innerk], 2 jl .nb232_checksingle jmp .nb232_unroll_loop.nb232_checksingle: mov edx, [esp + nb232_innerk] and edx, 1 jnz .nb232_dosingle jmp .nb232_updateouterdata.nb232_dosingle: mov edx, [esp + nb232_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx]
⌨️ 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?