nb_kernel214_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,167 行 · 第 1/5 页
S
2,167 行
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 + nb214_fixH2] movaps xmm1, [esp + nb214_fiyH2] movaps xmm2, [esp + nb214_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 ;# accumulate Mi forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb214_fixM] movaps xmm1, [esp + nb214_fiyM] movaps xmm2, [esp + nb214_fizM] 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 + 36] movss xmm4, [edi + ecx*4 + 40] movss xmm5, [edi + ecx*4 + 44] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [edi + ecx*4 + 36], xmm3 movss [edi + ecx*4 + 40], xmm4 movss [edi + ecx*4 + 44], 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 + nb214_n] ;# get group index for i particle mov edx, [ebp + nb214_gid] ;# base of gid[] mov edx, [edx + esi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movaps xmm7, [esp + nb214_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 + nb214_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 + nb214_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 + nb214_Vvdw] addss xmm7, [eax + edx*4] ;# move back to mem movss [eax + edx*4], xmm7 ;# finish if last mov ecx, [esp + nb214_nn1] ;# esi already loaded with n inc esi sub ecx, esi jecxz .nb214_outerend ;# not last, iterate outer loop once more! mov [esp + nb214_n], esi jmp .nb214_outer.nb214_outerend: ;# check if more outer neighborlists remain mov ecx, [esp + nb214_nri] ;# esi already loaded with n above sub ecx, esi jecxz .nb214_end ;# non-zero, do one more workunit jmp .nb214_threadloop.nb214_end: emms mov eax, [esp + nb214_nouter] mov ebx, [esp + nb214_ninner] mov ecx, [ebp + nb214_outeriter] mov edx, [ebp + nb214_inneriter] mov [ecx], eax mov [edx], ebx mov eax, [esp + nb214_salign] add esp, eax add esp, 1848 pop edi pop esi pop edx pop ecx pop ebx pop eax leave ret .globl nb_kernel214nf_ia32_sse.globl _nb_kernel214nf_ia32_ssenb_kernel214nf_ia32_sse: _nb_kernel214nf_ia32_sse: .equiv nb214nf_p_nri, 8.equiv nb214nf_iinr, 12.equiv nb214nf_jindex, 16.equiv nb214nf_jjnr, 20.equiv nb214nf_shift, 24.equiv nb214nf_shiftvec, 28.equiv nb214nf_fshift, 32.equiv nb214nf_gid, 36.equiv nb214nf_pos, 40.equiv nb214nf_faction, 44.equiv nb214nf_charge, 48.equiv nb214nf_p_facel, 52.equiv nb214nf_argkrf, 56.equiv nb214nf_argcrf, 60.equiv nb214nf_Vc, 64.equiv nb214nf_type, 68.equiv nb214nf_p_ntype, 72.equiv nb214nf_vdwparam, 76.equiv nb214nf_Vvdw, 80.equiv nb214nf_p_tabscale, 84.equiv nb214nf_VFtab, 88.equiv nb214nf_invsqrta, 92.equiv nb214nf_dvda, 96.equiv nb214nf_p_gbtabscale, 100.equiv nb214nf_GBtab, 104.equiv nb214nf_p_nthreads, 108.equiv nb214nf_count, 112.equiv nb214nf_mtx, 116.equiv nb214nf_outeriter, 120.equiv nb214nf_inneriter, 124.equiv nb214nf_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb214nf_ixO, 0.equiv nb214nf_iyO, 16.equiv nb214nf_izO, 32.equiv nb214nf_ixH1, 48.equiv nb214nf_iyH1, 64.equiv nb214nf_izH1, 80.equiv nb214nf_ixH2, 96.equiv nb214nf_iyH2, 112.equiv nb214nf_izH2, 128.equiv nb214nf_ixM, 144.equiv nb214nf_iyM, 160.equiv nb214nf_izM, 176.equiv nb214nf_jxO, 192.equiv nb214nf_jyO, 208.equiv nb214nf_jzO, 224.equiv nb214nf_jxH1, 240.equiv nb214nf_jyH1, 256.equiv nb214nf_jzH1, 272.equiv nb214nf_jxH2, 288.equiv nb214nf_jyH2, 304.equiv nb214nf_jzH2, 320.equiv nb214nf_jxM, 336.equiv nb214nf_jyM, 352.equiv nb214nf_jzM, 368.equiv nb214nf_qqMM, 384.equiv nb214nf_qqMH, 400.equiv nb214nf_qqHH, 416.equiv nb214nf_two, 432.equiv nb214nf_c6, 448.equiv nb214nf_c12, 464.equiv nb214nf_vctot, 480.equiv nb214nf_Vvdwtot, 496.equiv nb214nf_half, 512.equiv nb214nf_three, 528.equiv nb214nf_rsqOO, 544.equiv nb214nf_rsqH1H1, 560.equiv nb214nf_rsqH1H2, 576.equiv nb214nf_rsqH1M, 592.equiv nb214nf_rsqH2H1, 608.equiv nb214nf_rsqH2H2, 624.equiv nb214nf_rsqH2M, 640.equiv nb214nf_rsqMH1, 656.equiv nb214nf_rsqMH2, 672.equiv nb214nf_rsqMM, 688.equiv nb214nf_rinvsqOO, 704.equiv nb214nf_rinvH1H1, 720.equiv nb214nf_rinvH1H2, 736.equiv nb214nf_rinvH1M, 752.equiv nb214nf_rinvH2H1, 768.equiv nb214nf_rinvH2H2, 784.equiv nb214nf_rinvH2M, 800.equiv nb214nf_rinvMH1, 816.equiv nb214nf_rinvMH2, 832.equiv nb214nf_rinvMM, 848.equiv nb214nf_krf, 864.equiv nb214nf_crf, 880.equiv nb214nf_is3, 896.equiv nb214nf_ii3, 900.equiv nb214nf_innerjjnr, 904.equiv nb214nf_innerk, 908.equiv nb214nf_n, 912.equiv nb214nf_nn1, 916.equiv nb214nf_nri, 920.equiv nb214nf_nouter, 924.equiv nb214nf_ninner, 928.equiv nb214nf_salign, 932 push ebp mov ebp,esp push eax push ebx push ecx push edx push esi push edi sub esp, 936 ;# local stack space mov eax, esp and eax, 0xf sub esp, eax mov [esp + nb214nf_salign], eax emms ;# Move args passed by reference to stack mov ecx, [ebp + nb214nf_p_nri] mov ecx, [ecx] mov [esp + nb214nf_nri], ecx ;# zero iteration counters mov eax, 0 mov [esp + nb214nf_nouter], eax mov [esp + nb214nf_ninner], eax mov esi, [ebp + nb214nf_argkrf] mov edi, [ebp + nb214nf_argcrf] movss xmm5, [esi] movss xmm6, [edi] shufps xmm5, xmm5, 0 shufps xmm6, xmm6, 0 movaps [esp + nb214nf_krf], xmm5 movaps [esp + nb214nf_crf], xmm6 ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# constant 0.5 in IEEE (hex) mov [esp + nb214nf_half], eax movss xmm1, [esp + nb214nf_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 + nb214nf_half], xmm1 movaps [esp + nb214nf_two], xmm2 movaps [esp + nb214nf_three], xmm3 ;# assume we have at least one i particle - start directly mov ecx, [ebp + nb214nf_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx] ;# ebx =ii mov edx, [ebp + nb214nf_charge] movss xmm5, [edx + ebx*4 + 4] movss xmm3, [edx + ebx*4 + 12] movss xmm4, xmm3 mov esi, [ebp + nb214nf_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 + nb214nf_qqMM], xmm3 movaps [esp + nb214nf_qqMH], xmm4 movaps [esp + nb214nf_qqHH], xmm5 xorps xmm0, xmm0 mov edx, [ebp + nb214nf_type] mov ecx, [edx + ebx*4] shl ecx, 1 mov edx, ecx mov edi, [ebp + nb214nf_p_ntype] imul ecx, [edi] ;# ecx = ntia = 2*ntype*type[ii0] add edx, ecx mov eax, [ebp + nb214nf_vdwparam] movlps xmm0, [eax + edx*4] movaps xmm1, xmm0 shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 0x55 movaps [esp + nb214nf_c6], xmm0 movaps [esp + nb214nf_c12], xmm1.nb214nf_threadloop: mov esi, [ebp + nb214nf_count] ;# pointer to sync counter mov eax, [esi].nb214nf_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 .nb214nf_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [esp + nb214nf_nri] mov edx, ecx sub ecx, ebx cmovle ebx, edx ;# if(nn1>nri) nn1=nri ;# Cleared the spinlock if we got here. ;# eax contains nn0, ebx contains nn1. mov [esp + nb214nf_n], eax mov [esp + nb214nf_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb214nf_outerstart jmp .nb214nf_end .nb214nf_outerstart: ;# ebx contains number of outer iterations add ebx, [esp + nb214nf_nouter] mov [esp + nb214nf_nouter], ebx.nb214nf_outer: mov eax, [ebp + nb214nf_shift] ;# eax = pointer into shift[] mov ebx, [eax + esi*4] ;# ebx=shift[n] lea ebx, [ebx + ebx*2] ;# ebx=3*is mov [esp + nb214nf_is3],ebx ;# store is3 mov eax, [ebp + nb214nf_shiftvec] ;# eax = base of shiftvec[] movss xmm0, [eax + ebx*4] movss xmm1, [eax + ebx*4 + 4] movss xmm2, [eax + ebx*4 + 8] mov ecx, [ebp + nb214nf_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx + esi*4] ;# ebx =ii lea ebx, [ebx + ebx*2] ;# ebx = 3*ii=ii3 mov eax, [ebp + nb214nf_pos] ;# eax = base of pos[] mov [esp + nb214nf_ii3], ebx movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 movaps xmm6, xmm0 movaps xmm7, xmm1 addss xmm3, [eax + ebx*4] ;# ox addss xmm4⌨️ 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?