nb_kernel100_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,232 行 · 第 1/3 页
S
1,232 行
sub dword ptr [esp + nb100_innerk], 4 jl .nb100_finish_inner jmp .nb100_unroll_loop.nb100_finish_inner: ;# check if at least two particles remain add dword ptr [esp + nb100_innerk], 4 mov edx, [esp + nb100_innerk] and edx, 2 jnz .nb100_dopair jmp .nb100_checksingle.nb100_dopair: mov esi, [ebp + nb100_charge] mov edi, [ebp + nb100_pos] mov ecx, [esp + nb100_innerjjnr] mov eax, [ecx] mov ebx, [ecx + 4] add dword ptr [esp + nb100_innerjjnr], 8 movss xmm3, [esi + eax*4] movss xmm6, [esi + ebx*4] shufps xmm3, xmm6, 0 shufps xmm3, xmm3, 8 ;# constant 00001000 ;# xmm3(0,1) has the charges lea eax, [eax + eax*2] lea ebx, [ebx + ebx*2] ;# move coordinates to xmm0-xmm2 movlps xmm1, [edi + eax*4] movss xmm2, [edi + eax*4 + 8] movhps xmm1, [edi + ebx*4] movss xmm0, [edi + ebx*4 + 8] mulps xmm3, [esp + nb100_iq] xorps xmm7,xmm7 movlhps xmm3, xmm7 shufps xmm2, xmm0, 0 movaps xmm0, xmm1 shufps xmm2, xmm2, 136 ;# constant 10001000 shufps xmm0, xmm0, 136 ;# constant 10001000 shufps xmm1, xmm1, 221 ;# constant 11011101 mov edi, [ebp + nb100_faction] ;# move nb100_ix-iz to xmm4-xmm6 xorps xmm7, xmm7 movaps xmm4, [esp + nb100_ix] movaps xmm5, [esp + nb100_iy] movaps xmm6, [esp + nb100_iz] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb100_dx], xmm4 movaps [esp + nb100_dy], xmm5 movaps [esp + nb100_dz], xmm6 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm4, xmm5 addps xmm4, xmm6 ;# rsq in xmm4 rsqrtps xmm5, xmm4 ;# lookup seed in xmm5 movaps xmm2, xmm5 mulps xmm5, xmm5 movaps xmm1, [esp + nb100_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [esp + nb100_half] subps xmm1, xmm5 ;# constant 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv movaps xmm4, xmm0 mulps xmm4, xmm4 ;# xmm4=rinvsq movaps xmm5, [esp + nb100_vctot] mulps xmm3, xmm0 ;# xmm3=vcoul mulps xmm4, xmm3 ;# xmm4=fscal addps xmm5, xmm3 movaps xmm0, [esp + nb100_dx] movaps xmm1, [esp + nb100_dy] movaps xmm2, [esp + nb100_dz] movaps [esp + nb100_vctot], xmm5 mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# xmm0-xmm2 contains tx-tz (partial force) ;# now update f_i movaps xmm3, [esp + nb100_fix] movaps xmm4, [esp + nb100_fiy] movaps xmm5, [esp + nb100_fiz] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm5, xmm2 movaps [esp + nb100_fix], xmm3 movaps [esp + nb100_fiy], xmm4 movaps [esp + nb100_fiz], xmm5 ;# update the fj's movss xmm3, [edi + eax*4] movss xmm4, [edi + eax*4 + 4] movss xmm5, [edi + eax*4 + 8] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [edi + eax*4], xmm3 movss [edi + eax*4 + 4], xmm4 movss [edi + eax*4 + 8], xmm5 shufps xmm0, xmm0, 225 ;# constant 11100001 shufps xmm1, xmm1, 225 ;# constant 11100001 shufps xmm2, xmm2, 225 ;# constant 11100001 movss xmm3, [edi + ebx*4] movss xmm4, [edi + ebx*4 + 4] movss xmm5, [edi + ebx*4 + 8] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [edi + ebx*4], xmm3 movss [edi + ebx*4 + 4], xmm4 movss [edi + ebx*4 + 8], xmm5 .nb100_checksingle: mov edx, [esp + nb100_innerk] and edx, 1 jnz .nb100_dosingle jmp .nb100_updateouterdata.nb100_dosingle: mov esi, [ebp + nb100_charge] mov edi, [ebp + nb100_pos] mov ecx, [esp + nb100_innerjjnr] mov eax, [ecx] movss xmm3, [esi + eax*4] ;# xmm3(0) has the charge lea eax, [eax + eax*2] ;# move coordinates to xmm0-xmm2 movss xmm0, [edi + eax*4] movss xmm1, [edi + eax*4 + 4] movss xmm2, [edi + eax*4 + 8] mulps xmm3, [esp + nb100_iq] xorps xmm7, xmm7 movaps xmm4, [esp + nb100_ix] movaps xmm5, [esp + nb100_iy] movaps xmm6, [esp + nb100_iz] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb100_dx], xmm4 movaps [esp + nb100_dy], xmm5 movaps [esp + nb100_dz], xmm6 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm4, xmm5 addps xmm4, xmm6 ;# rsq in xmm4 rsqrtps xmm5, xmm4 ;# lookup seed in xmm5 movaps xmm2, xmm5 mulps xmm5, xmm5 movaps xmm1, [esp + nb100_three] mulps xmm5, xmm4 ;# rsq*lu*lu movaps xmm0, [esp + nb100_half] subps xmm1, xmm5 ;# constant 30-rsq*lu*lu mulps xmm1, xmm2 mulps xmm0, xmm1 ;# xmm0=rinv movaps xmm4, xmm0 mulps xmm4, xmm4 ;# xmm4=rinvsq mov edi, [ebp + nb100_faction] movaps xmm5, [esp + nb100_vctot] mulps xmm3, xmm0 ;# xmm3=vcoul mulps xmm4, xmm3 ;# xmm4=fscal addss xmm5, xmm3 movaps xmm0, [esp + nb100_dx] movaps xmm1, [esp + nb100_dy] movaps xmm2, [esp + nb100_dz] movaps [esp + nb100_vctot], xmm5 mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# xmm0-xmm2 contains tx-tz (partial force) ;# now update f_i movaps xmm3, [esp + nb100_fix] movaps xmm4, [esp + nb100_fiy] movaps xmm5, [esp + nb100_fiz] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movaps [esp + nb100_fix], xmm3 movaps [esp + nb100_fiy], xmm4 movaps [esp + nb100_fiz], xmm5 ;# update fj movss xmm3, [edi + eax*4] movss xmm4, [edi + eax*4 + 4] movss xmm5, [edi + eax*4 + 8] subss xmm3, xmm0 subss xmm4, xmm1 subss xmm5, xmm2 movss [edi + eax*4], xmm3 movss [edi + eax*4 + 4], xmm4 movss [edi + eax*4 + 8], xmm5 .nb100_updateouterdata: mov ecx, [esp + nb100_ii3] mov edi, [ebp + nb100_faction] mov esi, [ebp + nb100_fshift] mov edx, [esp + nb100_is3] ;# accumulate i forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb100_fix] movaps xmm1, [esp + nb100_fiy] movaps xmm2, [esp + nb100_fiz] 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 ;# increment fshift force movss xmm3, [esi + edx*4] movss xmm4, [esi + edx*4 + 4] movss xmm5, [esi + edx*4 + 8] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movss [esi + edx*4], xmm3 movss [esi + edx*4 + 4], xmm4 movss [esi + edx*4 + 8], xmm5 ;# get n from stack mov esi, [esp + nb100_n] ;# get group index for i particle mov edx, [ebp + nb100_gid] ;# base of gid[] mov edx, [edx + esi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movaps xmm7, [esp + nb100_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 + nb100_Vc] addss xmm7, [eax + edx*4] ;# move back to mem movss [eax + edx*4], xmm7 ;# finish if last mov ecx, [esp + nb100_nn1] ;# esi already loaded with n inc esi sub ecx, esi jecxz .nb100_outerend ;# not last, iterate outer loop once more! mov [esp + nb100_n], esi jmp .nb100_outer.nb100_outerend: ;# check if more outer neighborlists remain mov ecx, [esp + nb100_nri] ;# esi already loaded with n above sub ecx, esi jecxz .nb100_end ;# non-zero, do one more workunit jmp .nb100_threadloop.nb100_end: emms mov eax, [esp + nb100_nouter] mov ebx, [esp + nb100_ninner] mov ecx, [ebp + nb100_outeriter] mov edx, [ebp + nb100_inneriter] mov [ecx], eax mov [edx], ebx mov eax, [esp + nb100_salign] add esp, eax add esp, 252 pop edi pop esi pop edx pop ecx pop ebx pop eax leave ret.globl nb_kernel100nf_ia32_sse.globl _nb_kernel100nf_ia32_ssenb_kernel100nf_ia32_sse: _nb_kernel100nf_ia32_sse: .equiv nb100nf_p_nri, 8.equiv nb100nf_iinr, 12.equiv nb100nf_jindex, 16.equiv nb100nf_jjnr, 20.equiv nb100nf_shift, 24.equiv nb100nf_shiftvec, 28.equiv nb100nf_fshift, 32.equiv nb100nf_gid, 36.equiv nb100nf_pos, 40.equiv nb100nf_faction, 44.equiv nb100nf_charge, 48.equiv nb100nf_p_facel, 52.equiv nb100nf_p_krf, 56.equiv nb100nf_p_crf, 60.equiv nb100nf_Vc, 64.equiv nb100nf_type, 68.equiv nb100nf_p_ntype, 72.equiv nb100nf_vdwparam, 76.equiv nb100nf_Vvdw, 80.equiv nb100nf_p_tabscale, 84.equiv nb100nf_VFtab, 88.equiv nb100nf_invsqrta, 92.equiv nb100nf_dvda, 96.equiv nb100nf_p_gbtabscale, 100.equiv nb100nf_GBtab, 104.equiv nb100nf_p_nthreads, 108.equiv nb100nf_count, 112.equiv nb100nf_mtx, 116.equiv nb100nf_outeriter, 120.equiv nb100nf_inneriter, 124.equiv nb100nf_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb100nf_ix, 0.equiv nb100nf_iy, 16.equiv nb100nf_iz, 32.equiv nb100nf_iq, 48.equiv nb100nf_vctot, 64.equiv nb100nf_half, 80.equiv nb100nf_three, 96.equiv nb100nf_is3, 112.equiv nb100nf_ii3, 116.equiv nb100nf_innerjjnr, 120.equiv nb100nf_innerk, 124.equiv nb100nf_n, 128.equiv nb100nf_nn1, 132.equiv nb100nf_nri, 136.equiv nb100nf_facel, 140.equiv nb100nf_nouter, 144.equiv nb100nf_ninner, 148.equiv nb100nf_salign, 152 push ebp mov ebp,esp push eax push ebx push ecx push edx push esi push edi sub esp, 156 ;# local stack space mov eax, esp and eax, 0xf sub esp, eax mov [esp + nb100nf_salign], eax emms ;# Move args passed by reference to stack⌨️ 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?