nb_kernel112_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,177 行 · 第 1/5 页
S
2,177 行
subps xmm4, xmm1 subps xmm5, xmm2 mov esi, [ebp + nb112_faction] movaps [esp + nb112_fjxO], xmm3 movaps [esp + nb112_fjyO], xmm4 movaps [esp + nb112_fjzO], xmm5 addps xmm0, [esp + nb112_fixH2] addps xmm1, [esp + nb112_fiyH2] addps xmm2, [esp + nb112_fizH2] movaps [esp + nb112_fixH2], xmm0 movaps [esp + nb112_fiyH2], xmm1 movaps [esp + nb112_fizH2], xmm2 ;# update j water forces from local variables movlps xmm0, [esi + eax*4] movlps xmm1, [esi + eax*4 + 12] movhps xmm1, [esi + eax*4 + 24] movaps xmm3, [esp + nb112_fjxO] movaps xmm4, [esp + nb112_fjyO] movaps xmm5, [esp + nb112_fjzO] movaps xmm6, xmm5 movaps xmm7, xmm5 shufps xmm6, xmm6, 2 ;# constant 00000010 shufps xmm7, xmm7, 3 ;# constant 00000011 addss xmm5, [esi + eax*4 + 8] addss xmm6, [esi + eax*4 + 20] addss xmm7, [esi + eax*4 + 32] movss [esi + eax*4 + 8], xmm5 movss [esi + eax*4 + 20], xmm6 movss [esi + eax*4 + 32], xmm7 movaps xmm5, xmm3 unpcklps xmm3, xmm4 unpckhps xmm5, xmm4 addps xmm0, xmm3 addps xmm1, xmm5 movlps [esi + eax*4], xmm0 movlps [esi + eax*4 + 12], xmm1 movhps [esi + eax*4 + 24], xmm1 dec dword ptr [esp + nb112_innerk] jz .nb112_updateouterdata jmp .nb112_single_loop.nb112_updateouterdata: mov ecx, [esp + nb112_ii3] mov edi, [ebp + nb112_faction] mov esi, [ebp + nb112_fshift] mov edx, [esp + nb112_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb112_fixO] movaps xmm1, [esp + nb112_fiyO] movaps xmm2, [esp + nb112_fizO] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 1 shufps xmm4, xmm4, 1 shufps xmm5, xmm5, 1 addss xmm0, xmm3 addss xmm1, xmm4 addss xmm2, xmm5 ;# xmm0-xmm2 has single force in pos0 ;# increment i force movss xmm3, [edi + ecx*4] movss xmm4, [edi + ecx*4 + 4] movss xmm5, [edi + ecx*4 + 8] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [edi + ecx*4], xmm3 movss [edi + ecx*4 + 4], xmm4 movss [edi + ecx*4 + 8], xmm5 ;# accumulate force in xmm6/xmm7 for fshift movaps xmm6, xmm0 movss xmm7, xmm2 movlhps xmm6, xmm1 shufps xmm6, xmm6, 8 ;# constant 00001000 ;# accumulate H1i forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb112_fixH1] movaps xmm1, [esp + nb112_fiyH1] movaps xmm2, [esp + nb112_fizH1] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 1 shufps xmm4, xmm4, 1 shufps xmm5, xmm5, 1 addss xmm0, xmm3 addss xmm1, xmm4 addss xmm2, xmm5 ;# xmm0-xmm2 has single force in pos0 ;# increment i force movss xmm3, [edi + ecx*4 + 12] movss xmm4, [edi + ecx*4 + 16] movss xmm5, [edi + ecx*4 + 20] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [edi + ecx*4 + 12], xmm3 movss [edi + ecx*4 + 16], xmm4 movss [edi + ecx*4 + 20], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addss xmm7, xmm2 movlhps xmm0, xmm1 shufps xmm0, xmm0, 8 ;# constant 00001000 addps xmm6, xmm0 ;# accumulate H2i forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb112_fixH2] movaps xmm1, [esp + nb112_fiyH2] movaps xmm2, [esp + nb112_fizH2] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 ;# sum is in 1/2 in xmm0-xmm2 movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 1 shufps xmm4, xmm4, 1 shufps xmm5, xmm5, 1 addss xmm0, xmm3 addss xmm1, xmm4 addss xmm2, xmm5 ;# xmm0-xmm2 has single force in pos0 ;# increment i force movss xmm3, [edi + ecx*4 + 24] movss xmm4, [edi + ecx*4 + 28] movss xmm5, [edi + ecx*4 + 32] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [edi + ecx*4 + 24], xmm3 movss [edi + ecx*4 + 28], xmm4 movss [edi + ecx*4 + 32], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addss xmm7, xmm2 movlhps xmm0, xmm1 shufps xmm0, xmm0, 8 ;# constant 00001000 addps xmm6, xmm0 ;# increment fshift force movlps xmm3, [esi + edx*4] movss xmm4, [esi + edx*4 + 8] addps xmm3, xmm6 addss xmm4, xmm7 movlps [esi + edx*4], xmm3 movss [esi + edx*4 + 8], xmm4 ;# get n from stack mov esi, [esp + nb112_n] ;# get group index for i particle mov edx, [ebp + nb112_gid] ;# base of gid[] mov edx, [edx + esi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movaps xmm7, [esp + nb112_vctot] ;# accumulate movhlps xmm6, xmm7 addps xmm7, xmm6 ;# pos 0-1 in xmm7 have the sum now movaps xmm6, xmm7 shufps xmm6, xmm6, 1 addss xmm7, xmm6 ;# add earlier value from mem mov eax, [ebp + nb112_Vc] addss xmm7, [eax + edx*4] ;# move back to mem movss [eax + edx*4], xmm7 ;# accumulate total lj energy and update it movaps xmm7, [esp + nb112_Vvdwtot] ;# accumulate movhlps xmm6, xmm7 addps xmm7, xmm6 ;# pos 0-1 in xmm7 have the sum now movaps xmm6, xmm7 shufps xmm6, xmm6, 1 addss xmm7, xmm6 ;# add earlier value from mem mov eax, [ebp + nb112_Vvdw] addss xmm7, [eax + edx*4] ;# move back to mem movss [eax + edx*4], xmm7 ;# finish if last mov ecx, [esp + nb112_nn1] ;# esi already loaded with n inc esi sub ecx, esi jecxz .nb112_outerend ;# not last, iterate outer loop once more! mov [esp + nb112_n], esi jmp .nb112_outer.nb112_outerend: ;# check if more outer neighborlists remain mov ecx, [esp + nb112_nri] ;# esi already loaded with n above sub ecx, esi jecxz .nb112_end ;# non-zero, do one more workunit jmp .nb112_threadloop.nb112_end: emms mov eax, [esp + nb112_nouter] mov ebx, [esp + nb112_ninner] mov ecx, [ebp + nb112_outeriter] mov edx, [ebp + nb112_inneriter] mov [ecx], eax mov [edx], ebx mov eax, [esp + nb112_salign] add esp, eax add esp, 1512 pop edi pop esi pop edx pop ecx pop ebx pop eax leave ret .globl nb_kernel112nf_ia32_sse.globl _nb_kernel112nf_ia32_ssenb_kernel112nf_ia32_sse: _nb_kernel112nf_ia32_sse: .equiv nb112nf_p_nri, 8.equiv nb112nf_iinr, 12.equiv nb112nf_jindex, 16.equiv nb112nf_jjnr, 20.equiv nb112nf_shift, 24.equiv nb112nf_shiftvec, 28.equiv nb112nf_fshift, 32.equiv nb112nf_gid, 36.equiv nb112nf_pos, 40.equiv nb112nf_faction, 44.equiv nb112nf_charge, 48.equiv nb112nf_p_facel, 52.equiv nb112nf_krf, 56.equiv nb112nf_crf, 60.equiv nb112nf_Vc, 64.equiv nb112nf_type, 68.equiv nb112nf_p_ntype, 72.equiv nb112nf_vdwparam, 76.equiv nb112nf_Vvdw, 80.equiv nb112nf_p_tabscale, 84.equiv nb112nf_VFtab, 88.equiv nb112nf_invsqrta, 92.equiv nb112nf_dvda, 96.equiv nb112nf_p_gbtabscale, 100.equiv nb112nf_GBtab, 104.equiv nb112nf_p_nthreads, 108.equiv nb112nf_count, 112.equiv nb112nf_mtx, 116.equiv nb112nf_outeriter, 120.equiv nb112nf_inneriter, 124.equiv nb112nf_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb112nf_ixO, 0.equiv nb112nf_iyO, 16.equiv nb112nf_izO, 32.equiv nb112nf_ixH1, 48.equiv nb112nf_iyH1, 64.equiv nb112nf_izH1, 80.equiv nb112nf_ixH2, 96.equiv nb112nf_iyH2, 112.equiv nb112nf_izH2, 128.equiv nb112nf_jxO, 144.equiv nb112nf_jyO, 160.equiv nb112nf_jzO, 176.equiv nb112nf_jxH1, 192.equiv nb112nf_jyH1, 208.equiv nb112nf_jzH1, 224.equiv nb112nf_jxH2, 240.equiv nb112nf_jyH2, 256.equiv nb112nf_jzH2, 272.equiv nb112nf_qqOO, 288.equiv nb112nf_qqOH, 304.equiv nb112nf_qqHH, 320.equiv nb112nf_c6, 336.equiv nb112nf_c12, 352.equiv nb112nf_vctot, 368.equiv nb112nf_Vvdwtot, 384.equiv nb112nf_half, 400.equiv nb112nf_three, 416.equiv nb112nf_rsqOO, 432.equiv nb112nf_rsqOH1, 448.equiv nb112nf_rsqOH2, 464.equiv nb112nf_rsqH1O, 480.equiv nb112nf_rsqH1H1, 496.equiv nb112nf_rsqH1H2, 512.equiv nb112nf_rsqH2O, 528.equiv nb112nf_rsqH2H1, 544.equiv nb112nf_rsqH2H2, 560.equiv nb112nf_rinvOO, 576.equiv nb112nf_rinvOH1, 592.equiv nb112nf_rinvOH2, 608.equiv nb112nf_rinvH1O, 624.equiv nb112nf_rinvH1H1, 640.equiv nb112nf_rinvH1H2, 656.equiv nb112nf_rinvH2O, 672.equiv nb112nf_rinvH2H1, 688.equiv nb112nf_rinvH2H2, 704.equiv nb112nf_is3, 720.equiv nb112nf_ii3, 724.equiv nb112nf_innerjjnr, 728.equiv nb112nf_innerk, 732.equiv nb112nf_n, 736.equiv nb112nf_nn1, 740.equiv nb112nf_nri, 744.equiv nb112nf_nouter, 748.equiv nb112nf_ninner, 752.equiv nb112nf_salign, 756 push ebp mov ebp,esp push eax push ebx push ecx push edx push esi push edi sub esp, 760 ;# local stack space mov eax, esp and eax, 0xf sub esp, eax mov [esp + nb112nf_salign], eax emms ;# Move args passed by reference to stack mov ecx, [ebp + nb112nf_p_nri] mov ecx, [ecx] mov [esp + nb112nf_nri], ecx ;# zero iteration counters mov eax, 0 mov [esp + nb112nf_nouter], eax mov [esp + nb112nf_ninner], eax ;# assume we have at least one i particle - start directly mov ecx, [ebp + nb112nf_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx] ;# ebx =ii mov edx, [ebp + nb112nf_charge] movss xmm3, [edx + ebx*4] movss xmm4, xmm3 movss xmm5, [edx + ebx*4 + 4] mov esi, [ebp + nb112nf_p_facel] movss xmm6, [esi] mulss xmm3, xmm3 mulss xmm4, xmm5 mulss xmm5, xmm5 mulss xmm3, xmm6 mulss xmm4, xmm6 mulss xmm5, xmm6 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [esp + nb112nf_qqOO], xmm3 movaps [esp + nb112nf_qqOH], xmm4 movaps [esp + nb112nf_qqHH], xmm5 xorps xmm0, xmm0 mov edx, [ebp + nb112nf_type] mov ecx, [edx + ebx*4] shl ecx, 1 mov edx, ecx mov edi, [ebp + nb112nf_p_ntype] imul ecx, [edi] ;# ecx = ntia = 2*ntype*type[ii0] add edx, ecx mov eax, [ebp + nb112nf_vdwparam] movlps xmm0, [eax + edx*4] movaps xmm1, xmm0 shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 85 ;# constant 01010101 movaps [esp + nb112nf_c6], xmm0 movaps [esp + nb112nf_c12], xmm1 ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# constant 0.5 in IEEE (hex) mov [esp + nb112nf_half], eax movss xmm1, [esp + nb112nf_half] shufps xmm1, xmm1, 0 ;# splat to all elements movaps xmm2, xmm1 addps xmm2, xmm2 ;# constant 1.0 movaps xmm3, xmm2 addps xmm2, xmm2 ;# constant 2.0 addps xmm3, xmm2 ;# constant 3.0 movaps [esp + nb112nf_half], xmm1 movaps [esp + nb112nf_three], xmm3.nb112nf_threadloop: mov esi, [ebp + nb112nf_count] ;# pointer to sync counter mov eax, [esi].nb112nf_spinlock: mov ebx, eax ;# ebx=*count=nn0 add ebx, 1 ;# ebx=nn1=nn0+10 lock cmpxchg [esi], ebx ;# write nn1 to *counter, ;# if it hasnt changed. ;# or reread *counter to eax. pause ;# -> better p4 performance jnz .nb112nf_spinlock⌨️ 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?