nb_kernel101_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,492 行 · 第 1/3 页
S
1,492 行
movaps xmm4, xmm6 mulps xmm4, xmm4 ;# xmm6=rinv, xmm4=rinvsq mulps xmm6, [esp + nb101_qqH] ;# xmm6=vcoul mulps xmm4, xmm6 ;# total fsH1 in xmm4 addps xmm6, [esp + nb101_vctot] movaps xmm0, [esp + nb101_dxH1] movaps xmm1, [esp + nb101_dyH1] movaps xmm2, [esp + nb101_dzH1] movaps [esp + nb101_vctot], xmm6 mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# update H1 forces movaps xmm3, [esp + nb101_fixH1] movaps xmm4, [esp + nb101_fiyH1] movaps xmm7, [esp + nb101_fizH1] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm7, xmm2 movaps [esp + nb101_fixH1], xmm3 movaps [esp + nb101_fiyH1], xmm4 movaps [esp + nb101_fizH1], xmm7 ;# update j forces with water H1 addps xmm0, [esp + nb101_fjx] addps xmm1, [esp + nb101_fjy] addps xmm2, [esp + nb101_fjz] movaps [esp + nb101_fjx], xmm0 movaps [esp + nb101_fjy], xmm1 movaps [esp + nb101_fjz], xmm2 ;# H2 interactions movaps xmm4, xmm5 mulps xmm4, xmm4 ;# xmm5=rinv, xmm4=rinvsq mulps xmm5, [esp + nb101_qqH] ;# xmm5=vcoul mulps xmm4, xmm5 ;# total fsH1 in xmm4 addps xmm5, [esp + nb101_vctot] movaps xmm0, [esp + nb101_dxH2] movaps xmm1, [esp + nb101_dyH2] movaps xmm2, [esp + nb101_dzH2] movaps [esp + nb101_vctot], xmm5 mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# update H2 forces movaps xmm3, [esp + nb101_fixH2] movaps xmm4, [esp + nb101_fiyH2] movaps xmm7, [esp + nb101_fizH2] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm7, xmm2 movaps [esp + nb101_fixH2], xmm3 movaps [esp + nb101_fiyH2], xmm4 movaps [esp + nb101_fizH2], xmm7 mov edi, [ebp + nb101_faction] ;# update j forces addps xmm0, [esp + nb101_fjx] addps xmm1, [esp + nb101_fjy] addps xmm2, [esp + nb101_fjz] movlps xmm4, [edi + eax*4] movlps xmm7, [edi + ecx*4] movhps xmm4, [edi + ebx*4] movhps xmm7, [edi + edx*4] movaps xmm3, xmm4 shufps xmm3, xmm7, 136 ;# constant 10001000 shufps xmm4, xmm7, 221 ;# constant 11011101 ;# xmm3 has fjx, xmm4 has fjy subps xmm3, xmm0 subps xmm4, xmm1 ;# unpack them back for storing movaps xmm7, xmm3 unpcklps xmm7, xmm4 unpckhps xmm3, xmm4 movlps [edi + eax*4], xmm7 movlps [edi + ecx*4], xmm3 movhps [edi + ebx*4], xmm7 movhps [edi + edx*4], xmm3 ;# finally z forces movss xmm0, [edi + eax*4 + 8] movss xmm1, [edi + ebx*4 + 8] movss xmm3, [edi + ecx*4 + 8] movss xmm4, [edi + edx*4 + 8] subss xmm0, xmm2 shufps xmm2, xmm2, 229 ;# constant 11100101 subss xmm1, xmm2 shufps xmm2, xmm2, 234 ;# constant 11101010 subss xmm3, xmm2 shufps xmm2, xmm2, 255 ;# constant 11111111 subss xmm4, xmm2 movss [edi + eax*4 + 8], xmm0 movss [edi + ebx*4 + 8], xmm1 movss [edi + ecx*4 + 8], xmm3 movss [edi + edx*4 + 8], xmm4 ;# should we do one more iteration? sub dword ptr [esp + nb101_innerk], 4 jl .nb101_odd_inner jmp .nb101_unroll_loop.nb101_odd_inner: add dword ptr [esp + nb101_innerk], 4 jnz .nb101_odd_loop jmp .nb101_updateouterdata.nb101_odd_loop: mov edx, [esp + nb101_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] add dword ptr [esp + nb101_innerjjnr], 4 xorps xmm4, xmm4 movss xmm4, [esp + nb101_iqO] mov esi, [ebp + nb101_charge] movhps xmm4, [esp + nb101_iqH] movss xmm3, [esi + eax*4] ;# charge in xmm3 shufps xmm3, xmm3, 0 mulps xmm3, xmm4 movaps [esp + nb101_qqO], xmm3 ;# use oxygen qq for storage mov esi, [ebp + nb101_pos] lea eax, [eax + eax*2] ;# move j coords to xmm0-xmm2 movss xmm0, [esi + eax*4] movss xmm1, [esi + eax*4 + 4] movss xmm2, [esi + eax*4 + 8] shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 0 shufps xmm2, xmm2, 0 movss xmm3, [esp + nb101_ixO] movss xmm4, [esp + nb101_iyO] movss xmm5, [esp + nb101_izO] movlps xmm6, [esp + nb101_ixH1] movlps xmm7, [esp + nb101_ixH2] unpcklps xmm6, xmm7 movlhps xmm3, xmm6 movlps xmm6, [esp + nb101_iyH1] movlps xmm7, [esp + nb101_iyH2] unpcklps xmm6, xmm7 movlhps xmm4, xmm6 movlps xmm6, [esp + nb101_izH1] movlps xmm7, [esp + nb101_izH2] unpcklps xmm6, xmm7 movlhps xmm5, xmm6 subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 movaps [esp + nb101_dxO], xmm3 movaps [esp + nb101_dyO], xmm4 movaps [esp + nb101_dzO], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm4, xmm3 addps xmm4, xmm5 ;# rsq in xmm4 rsqrtps xmm5, xmm4 ;# lookup seed in xmm5 movaps xmm2, xmm5 mulps xmm5, xmm5 movaps xmm1, [esp + nb101_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [esp + nb101_half] subps xmm1, xmm5 ;# constant 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv ;# a little trick to avoid NaNs: ;# positions 0,2,and 3 are valid, but not 1. ;# If it contains NaN it doesnt help to mult by 0, ;# So we shuffle it and copy pos 0 to pos1! shufps xmm0, xmm0, 224 ;# constant 11100000 movaps xmm4, xmm0 mulps xmm4, xmm4 ;# xmm4=rinvsq movaps xmm3, [esp + nb101_qqO] mulps xmm3, xmm0 ;# xmm3=vcoul mulps xmm4, xmm3 ;# xmm4=total fscal addps xmm3, [esp + nb101_vctot] movaps xmm0, [esp + nb101_dxO] movaps xmm1, [esp + nb101_dyO] movaps xmm2, [esp + nb101_dzO] movaps [esp + nb101_vctot], xmm3 mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# xmm0-xmm2 contains tx-tz (partial force) movss xmm3, [esp + nb101_fixO] movss xmm4, [esp + nb101_fiyO] movss xmm5, [esp + nb101_fizO] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [esp + nb101_fixO], xmm3 movss [esp + nb101_fiyO], xmm4 movss [esp + nb101_fizO], xmm5 ;# updated the O force now do the H's movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 shufps xmm3, xmm3, 230 ;# constant 11100110 ;# shift right shufps xmm4, xmm4, 230 ;# constant 11100110 shufps xmm5, xmm5, 230 ;# constant 11100110 addss xmm3, [esp + nb101_fixH1] addss xmm4, [esp + nb101_fiyH1] addss xmm5, [esp + nb101_fizH1] movss [esp + nb101_fixH1], xmm3 movss [esp + nb101_fiyH1], xmm4 movss [esp + nb101_fizH1], xmm5 ;# updated the H1 force mov edi, [ebp + nb101_faction] shufps xmm3, xmm3, 231 ;# constant 11100111 ;# shift right shufps xmm4, xmm4, 231 ;# constant 11100111 shufps xmm5, xmm5, 231 ;# constant 11100111 addss xmm3, [esp + nb101_fixH2] addss xmm4, [esp + nb101_fiyH2] addss xmm5, [esp + nb101_fizH2] movss [esp + nb101_fixH2], xmm3 movss [esp + nb101_fiyH2], xmm4 movss [esp + nb101_fizH2], xmm5 ;# updated the H2 force ;# the fj's - start by accumulating the tx/ty/tz force in xmm0, xmm1 xorps xmm5, xmm5 movaps xmm3, xmm0 movlps xmm6, [edi + eax*4] movss xmm7, [edi + eax*4 + 8] unpcklps xmm3, xmm1 movlhps xmm3, xmm5 unpckhps xmm0, xmm1 addps xmm0, xmm3 movhlps xmm3, xmm0 addps xmm0, xmm3 ;# x,y sum in xmm0 movhlps xmm1, xmm2 addss xmm2, xmm1 shufps xmm1, xmm1, 1 addss xmm2, xmm1 ;# z sum in xmm2 subps xmm6, xmm0 subss xmm7, xmm2 movlps [edi + eax*4], xmm6 movss [edi + eax*4 + 8], xmm7 dec dword ptr [esp + nb101_innerk] jz .nb101_updateouterdata jmp .nb101_odd_loop.nb101_updateouterdata: mov ecx, [esp + nb101_ii3] mov edi, [ebp + nb101_faction] mov esi, [ebp + nb101_fshift] mov edx, [esp + nb101_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb101_fixO] movaps xmm1, [esp + nb101_fiyO] movaps xmm2, [esp + nb101_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 + nb101_fixH1] movaps xmm1, [esp + nb101_fiyH1] movaps xmm2, [esp + nb101_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 + nb101_fixH2] movaps xmm1, [esp + nb101_fiyH2] movaps xmm2, [esp + nb101_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 + nb101_n] ;# get group index for i particle mov edx, [ebp + nb101_gid] ;# base of gid[] mov edx, [edx + esi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movaps xmm7, [esp + nb101_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 + nb101_Vc] addss xmm7, [eax + edx*4] ;# move back to mem movss [eax + edx*4], xmm7 ;# finish if last mov ecx, [esp + nb101_nn1] ;# esi already loaded with n inc esi sub ecx, esi jecxz .nb101_outerend ;# not last, iterate outer loop once more! mov [esp + nb101_n], esi jmp .nb101_outer.nb101_outerend: ;# check if more outer neighborlists remain mov ecx, [esp + nb101_nri] ;# esi already loaded with n above sub ecx, esi jecxz .nb101_end ;# non-zero, do one more workunit jmp .nb101_threadloop.nb101_end: emms mov eax, [esp + nb101_nouter] mov ebx, [esp + nb101_ninner] mov ecx, [ebp + nb101_outeriter] mov edx, [ebp + nb101_inneriter] mov [ecx], eax mov [edx], ebx mov eax, [esp + nb101_salign] add esp, eax add esp, 632 pop edi pop esi pop edx pop ecx pop ebx pop eax leave ret.globl nb_kernel101nf_ia32_sse.globl _nb_kernel101nf_ia32_ssenb_kernel101nf_ia32_sse: _nb_kernel101nf_ia32_sse: .equiv nb101nf_p_nri, 8.equiv nb101nf_iinr, 12.equiv nb101nf_jindex, 16.equiv nb101nf_jjnr, 20.equiv nb101nf_shift, 24.equiv nb101nf_shiftvec, 28.equiv nb101nf_fshift, 32.equiv nb101nf_gid, 36.equiv nb101nf_pos, 40.equiv nb101nf_faction, 44.equiv nb101nf_charge, 48.equiv nb101nf_p_facel, 52.equiv nb101nf_p_krf, 56.equiv nb101nf_p_crf, 60.equiv nb101nf_Vc, 64.equiv nb101nf_type, 68.equiv nb101nf_p_ntype, 72.equiv nb101nf_vdwparam, 76.equiv nb101nf_Vvdw, 80.equiv nb101nf_p_tabscale, 84.equiv nb101nf_VFtab, 88.equiv nb101nf_invsqrta, 92.equiv nb101nf_dvda, 96.equiv nb101nf_p_gbtabscale, 100.equiv nb101nf_GBtab, 104.equiv nb101nf_p_nthreads, 108.equiv nb101nf_count, 112.equiv nb101nf_mtx, 116.equiv nb101nf_outeriter, 120.equiv nb101nf_inneriter, 124.equiv nb101nf_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb101nf_ixO, 0.equiv nb101nf_iyO, 16.equiv nb101nf_izO, 32.equiv nb101nf_ixH1, 48.equiv nb101nf_iyH1, 64⌨️ 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?