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
movlpd [esp + nb204_fjzH1], xmm5 movlpd [esp + nb204_fixH1], xmm0 movlpd [esp + nb204_fiyH1], xmm1 movlpd [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 mulsd xmm5, [esp + nb204_rsqH1H2] ;# xmm5=krsq movapd xmm4, xmm5 addsd xmm4, xmm7 ;# xmm4=r inv+ krsq mulsd xmm0, xmm0 subsd xmm4, [esp + nb204_crf] mulsd xmm4, [esp + nb204_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulsd xmm5, [esp + nb204_two] subsd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulsd xmm7, [esp + nb204_qqHH] ;# xmm7 = coul part of fscal addsd xmm6, xmm4 ;# add to local vctot mulsd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxH2] movapd xmm4, [esp + nb204_fjyH2] movapd xmm5, [esp + nb204_fjzH2] mulsd xmm0, [esp + nb204_dxH1H2] mulsd xmm1, [esp + nb204_dyH1H2] mulsd xmm2, [esp + nb204_dzH1H2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb204_fixH1] addsd xmm1, [esp + nb204_fiyH1] addsd xmm2, [esp + nb204_fizH1] movlpd [esp + nb204_fjxH2], xmm3 movlpd [esp + nb204_fjyH2], xmm4 movlpd [esp + nb204_fjzH2], xmm5 movlpd [esp + nb204_fixH1], xmm0 movlpd [esp + nb204_fiyH1], xmm1 movlpd [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 mulsd xmm5, [esp + nb204_rsqH2M] ;# xmm5=krsq movapd xmm4, xmm5 addsd xmm4, xmm7 ;# xmm4=r inv+ krsq subsd xmm4, [esp + nb204_crf] mulsd xmm0, xmm0 mulsd xmm4, [esp + nb204_qqMH] ;# xmm4=voul=qq*(rinv+ krsq) mulsd xmm5, [esp + nb204_two] subsd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulsd xmm7, [esp + nb204_qqMH] ;# xmm7 = coul part of fscal addsd xmm6, xmm4 ;# add to local vctot mulsd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxM] movapd xmm4, [esp + nb204_fjyM] movapd xmm5, [esp + nb204_fjzM] mulsd xmm0, [esp + nb204_dxH2M] mulsd xmm1, [esp + nb204_dyH2M] mulsd xmm2, [esp + nb204_dzH2M] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb204_fixH2] addsd xmm1, [esp + nb204_fiyH2] addsd xmm2, [esp + nb204_fizH2] movlpd [esp + nb204_fjxM], xmm3 movlpd [esp + nb204_fjyM], xmm4 movlpd [esp + nb204_fjzM], xmm5 movlpd [esp + nb204_fixH2], xmm0 movlpd [esp + nb204_fiyH2], xmm1 movlpd [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 mulsd xmm5, [esp + nb204_rsqH2H1] ;# xmm5=krsq movapd xmm4, xmm5 addsd xmm4, xmm7 ;# xmm4=r inv+ krsq subsd xmm4, [esp + nb204_crf] mulsd xmm0, xmm0 mulsd xmm4, [esp + nb204_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulsd xmm5, [esp + nb204_two] subsd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulsd xmm7, [esp + nb204_qqHH] ;# xmm7 = coul part of fscal addsd xmm6, xmm4 ;# add to local vctot mulsd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxH1] movapd xmm4, [esp + nb204_fjyH1] movapd xmm5, [esp + nb204_fjzH1] mulsd xmm0, [esp + nb204_dxH2H1] mulsd xmm1, [esp + nb204_dyH2H1] mulsd xmm2, [esp + nb204_dzH2H1] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb204_fixH2] addsd xmm1, [esp + nb204_fiyH2] addsd xmm2, [esp + nb204_fizH2] movlpd [esp + nb204_fjxH1], xmm3 movlpd [esp + nb204_fjyH1], xmm4 movlpd [esp + nb204_fjzH1], xmm5 movlpd [esp + nb204_fixH2], xmm0 movlpd [esp + nb204_fiyH2], xmm1 movlpd [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 mulsd xmm5, [esp + nb204_rsqH2H2] ;# xmm5=krsq movapd xmm4, xmm5 addsd xmm4, xmm7 ;# xmm4=r inv+ krsq subsd xmm4, [esp + nb204_crf] mulsd xmm0, xmm0 mulsd xmm4, [esp + nb204_qqHH] ;# xmm4=voul=qq*(rinv+ krsq) mulsd xmm5, [esp + nb204_two] subsd xmm7, xmm5 ;# xmm7=rinv-2*krsq mulsd xmm7, [esp + nb204_qqHH] ;# xmm7 = coul part of fscal addsd xmm6, xmm4 ;# add to local vctot mulsd xmm0, xmm7 ;# fsMH2 movapd xmm1, xmm0 movapd xmm2, xmm0 movapd xmm1, xmm0 movlpd [esp + nb204_vctot], xmm6 movapd xmm2, xmm0 movapd xmm3, [esp + nb204_fjxH2] movapd xmm4, [esp + nb204_fjyH2] movapd xmm5, [esp + nb204_fjzH2] mulsd xmm0, [esp + nb204_dxH2H2] mulsd xmm1, [esp + nb204_dyH2H2] mulsd xmm2, [esp + nb204_dzH2H2] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 addsd xmm0, [esp + nb204_fixH2] addsd xmm1, [esp + nb204_fiyH2] addsd xmm2, [esp + nb204_fizH2] movlpd [esp + nb204_fjxH2], xmm3 movlpd [esp + nb204_fjyH2], xmm4 movlpd [esp + nb204_fjzH2], xmm5 movlpd [esp + nb204_fixH2], xmm0 movlpd [esp + nb204_fiyH2], xmm1 movlpd [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] addsd xmm0, [esp + nb204_fjxH1] addsd xmm1, [esp + nb204_fjyH1] addsd xmm2, [esp + nb204_fjzH1] addsd xmm3, [esp + nb204_fjxH2] addsd xmm4, [esp + nb204_fjyH2] addsd xmm5, [esp + nb204_fjzH2] addsd xmm6, [esp + nb204_fjxM] addsd 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 movlpd xmm0, [edi + eax*8 + 88] addsd xmm0, [esp + nb204_fjzM] movlpd [edi + eax*8 + 88], xmm0 .nb204_updateouterdata: mov ecx, [esp + nb204_ii3] mov edi, [ebp + nb204_faction] mov esi, [ebp + nb204_fshift] mov edx, [esp + nb204_is3] ;# accumulate H1i forces in xmm0, xmm1, xmm2 movapd xmm0, [esp + nb204_fixH1] movapd xmm1, [esp + nb204_fiyH1] movapd xmm2, [esp + nb204_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 movapd xmm6, xmm0 movsd xmm7, xmm2 unpcklpd xmm6, xmm1 ;# accumulate H2i forces in xmm0, xmm1, xmm2 movapd xmm0, [esp + nb204_fixH2] movapd xmm1, [esp + nb204_fiyH2] movapd xmm2, [esp + nb204_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 ;# 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 H2i forces in xmm0, xmm1, xmm2 movapd xmm0, [esp + nb204_fixM] movapd xmm1, [esp + nb204_fiyM] movapd xmm2, [esp + nb204_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 + nb204_n] ;# get group index for i particle mov edx, [ebp + nb204_gid] ;# base of gid[] mov edx, [edx + esi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movapd xmm7, [esp + nb204_vctot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# low xmm7 has the sum now ;# add earlier value from mem mov eax, [ebp + nb204_Vc] addsd xmm7, [eax + edx*8] ;# move back to mem movsd [eax + edx*8], xmm7 ;# finish if last mov ecx, [esp + nb204_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb204_outerend ;# not last, iterate outer loop once more! mov [esp + nb204_n], esi jmp .nb204_outer.nb204_outerend: ;# check if more outer neighborlists remain mov ecx, [esp + nb204_nri] ;# esi already loaded with n above sub ecx, esi jz .nb204_end ;# non-zero, do one more workunit jmp .nb204_threadloop.nb204_end: emms mov eax, [esp + nb204_nouter] mov ebx, [esp + nb204_ninner] mov ecx, [ebp + nb204_outeriter] mov edx, [ebp + nb204_inneriter] mov [ecx], eax mov [edx], ebx mov eax, [esp + nb204_salign] add esp, eax add esp, 1480 pop edi pop esi pop edx pop ecx pop ebx pop eax leave ret .globl nb_kernel204nf_ia32_sse2.globl _nb_kernel204nf_ia32_sse2nb_kernel204nf_ia32_sse2: _nb_kernel204nf_ia32_sse2: .equiv nb204nf_p_nri, 8.equiv nb204nf_iinr, 12.equiv nb204nf_jindex, 16.equiv nb204nf_jjnr, 20.equiv nb204nf_shift, 24.equiv nb204nf_shiftvec, 28.equiv nb204nf_fshift, 32.equiv nb204nf_gid, 36.equiv nb204nf_pos, 40.equiv nb204nf_faction, 44.equiv nb204nf_charge, 48.equiv nb204nf_p_facel, 52.equiv nb204nf_argkrf, 56.equiv nb204nf_argcrf, 60.equiv nb204nf_Vc, 64.equiv nb204nf_type, 68.equiv nb204nf_p_ntype, 72.equiv nb204nf_vdwparam, 76.equiv nb204nf_Vvdw, 80.equiv nb204nf_p_tabscale, 84.equiv nb204nf_VFtab, 88.equiv nb204nf_invsqrta, 92.equiv nb204nf_dvda, 96.equiv nb204nf_p_gbtabscale, 100.equiv nb204nf_GBtab, 104.equiv nb204nf_p_nthreads, 108.equiv nb204nf_count, 112.equiv nb204nf_mtx, 116.equiv nb204nf_outeriter, 120.equiv nb204nf_inneriter, 124.equiv nb204nf_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb204nf_ixM, 0.equiv nb204nf_iyM, 16.equiv nb204nf_izM, 32.equiv nb204nf_ixH1, 48.equiv nb204nf_iyH1, 64.equiv nb204nf_izH1, 80.equiv nb204nf_ixH2, 96.equiv nb204nf_iyH2, 112.equiv nb204nf_izH2, 128.equiv nb204nf_jxM, 144.equiv nb204nf_jyM, 160.equiv nb204nf_jzM, 176.equiv nb204nf_jxH1, 192.equiv nb204nf_jyH1, 208.equiv nb204nf_jzH1, 224.equiv nb204nf_jxH2, 240.equiv nb204nf_jyH2, 256.equiv nb204nf_jzH2, 272.equiv nb204nf_qqMM, 288.equiv nb204nf_qqMH, 304.equiv nb204nf_qqHH, 320.equiv nb204nf_vctot, 336.equiv nb204nf_half, 352.equiv nb204nf_three, 368.equiv nb204nf_rsqMM, 384.equiv nb204nf_rsqMH1, 400.equiv nb204nf_rsqMH2, 416.equiv nb204nf_rsqH1M, 432.equiv nb204nf_rsqH1H1, 448.equiv nb2⌨️ 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?