nb_kernel010_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,380 行 · 第 1/3 页
S
1,380 行
.nb010_dopair: mov ecx, [esp + nb010_innerjjnr] mov eax, [ecx] mov ebx, [ecx + 4] add dword ptr [esp + nb010_innerjjnr], 8 mov esi, [ebp + nb010_type] mov ecx, eax mov edx, ebx mov ecx, [esi + ecx*4] mov edx, [esi + edx*4] mov esi, [ebp + nb010_vdwparam] shl ecx, 1 shl edx, 1 mov edi, [esp + nb010_ntia] add ecx, edi add edx, edi movlps xmm6, [esi + ecx*4] movhps xmm6, [esi + edx*4] mov edi, [ebp + nb010_pos] xorps xmm7,xmm7 movaps xmm4, xmm6 shufps xmm4, xmm4, 8 ;# constant 00001000 shufps xmm6, xmm6, 13 ;# constant 00001101 movlhps xmm4, xmm7 movlhps xmm6, xmm7 movaps [esp + nb010_c6], xmm4 movaps [esp + nb010_c12], xmm6 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] 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 + nb010_faction] ;# move nb010_ix-iz to xmm4-xmm6 xorps xmm7, xmm7 movaps xmm4, [esp + nb010_ix] movaps xmm5, [esp + nb010_iy] movaps xmm6, [esp + nb010_iz] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb010_dx], xmm4 movaps [esp + nb010_dy], xmm5 movaps [esp + nb010_dz], xmm6 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm4, xmm5 addps xmm4, xmm6 ;# rsq in xmm4 rcpps xmm5, xmm4 ;# constant 1/x lookup seed in xmm5 movaps xmm0, [esp + nb010_two] mulps xmm4, xmm5 subps xmm0, xmm4 mulps xmm0, xmm5 ;# xmm0=rinvsq movaps xmm4, xmm0 movaps xmm1, xmm0 mulps xmm1, xmm0 mulps xmm1, xmm0 ;# xmm1=rinvsix movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=rinvtwelve mulps xmm1, [esp + nb010_c6] mulps xmm2, [esp + nb010_c12] movaps xmm5, xmm2 subps xmm5, xmm1 ;# Vvdw=Vvdw12-Vvdw6 addps xmm5, [esp + nb010_Vvdwtot] mulps xmm1, [esp + nb010_six] mulps xmm2, [esp + nb010_twelve] subps xmm2, xmm1 mulps xmm4, xmm2 ;# xmm4=total fscal movaps xmm0, [esp + nb010_dx] movaps xmm1, [esp + nb010_dy] movaps xmm2, [esp + nb010_dz] movaps [esp + nb010_Vvdwtot], xmm5 mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# xmm0-xmm2 contains tx-tz (partial force) ;# now update f_i movaps xmm3, [esp + nb010_fix] movaps xmm4, [esp + nb010_fiy] movaps xmm5, [esp + nb010_fiz] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm5, xmm2 movaps [esp + nb010_fix], xmm3 movaps [esp + nb010_fiy], xmm4 movaps [esp + nb010_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 .nb010_checksingle: mov edx, [esp + nb010_innerk] and edx, 1 jnz .nb010_dosingle jmp .nb010_updateouterdata.nb010_dosingle: mov edi, [ebp + nb010_pos] mov ecx, [esp + nb010_innerjjnr] mov eax, [ecx] mov esi, [ebp + nb010_type] mov ecx, eax mov ecx, [esi + ecx*4] mov esi, [ebp + nb010_vdwparam] shl ecx, 1 add ecx, [esp + nb010_ntia] xorps xmm6, xmm6 movlps xmm6, [esi + ecx*4] movaps xmm4, xmm6 shufps xmm4, xmm4, 252 ;# constant 11111100 shufps xmm6, xmm6, 253 ;# constant 11111101 movaps [esp + nb010_c6], xmm4 movaps [esp + nb010_c12], xmm6 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] xorps xmm7, xmm7 movaps xmm4, [esp + nb010_ix] movaps xmm5, [esp + nb010_iy] movaps xmm6, [esp + nb010_iz] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb010_dx], xmm4 movaps [esp + nb010_dy], xmm5 movaps [esp + nb010_dz], xmm6 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm4, xmm5 addps xmm4, xmm6 ;# rsq in xmm4 rcpps xmm5, xmm4 ;# constant 1/x lookup seed in xmm5 movaps xmm0, [esp + nb010_two] mulps xmm4, xmm5 subps xmm0, xmm4 mulps xmm0, xmm5 ;# xmm0=rinvsq movaps xmm4, xmm0 movaps xmm1, xmm0 mulps xmm1, xmm0 mulps xmm1, xmm0 ;# xmm1=rinvsix movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=rinvtwelve mulps xmm1, [esp + nb010_c6] mulps xmm2, [esp + nb010_c12] movaps xmm5, xmm2 subps xmm5, xmm1 ;# Vvdw=Vvdw12-Vvdw6 addss xmm5, [esp + nb010_Vvdwtot] mulps xmm1, [esp + nb010_six] mulps xmm2, [esp + nb010_twelve] subps xmm2, xmm1 mulps xmm4, xmm2 ;# xmm4=total fscal mov edi, [ebp + nb010_faction] movaps xmm0, [esp + nb010_dx] movaps xmm1, [esp + nb010_dy] movaps xmm2, [esp + nb010_dz] movss [esp + nb010_Vvdwtot], xmm5 mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# xmm0-xmm2 contains tx-tz (partial force) ;# now update f_i movaps xmm3, [esp + nb010_fix] movaps xmm4, [esp + nb010_fiy] movaps xmm5, [esp + nb010_fiz] addss xmm3, xmm0 addss xmm4, xmm1 addss xmm5, xmm2 movaps [esp + nb010_fix], xmm3 movaps [esp + nb010_fiy], xmm4 movaps [esp + nb010_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 .nb010_updateouterdata: mov ecx, [esp + nb010_ii3] mov edi, [ebp + nb010_faction] mov esi, [ebp + nb010_fshift] mov edx, [esp + nb010_is3] ;# accumulate i forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb010_fix] movaps xmm1, [esp + nb010_fiy] movaps xmm2, [esp + nb010_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 + nb010_n] ;# get group index for i particle mov edx, [ebp + nb010_gid] ;# base of gid[] mov edx, [edx + esi*4] ;# ggid=gid[n] ;# accumulate total lj energy and update it movaps xmm7, [esp + nb010_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 + nb010_Vvdw] addss xmm7, [eax + edx*4] ;# move back to mem movss [eax + edx*4], xmm7 ;# finish if last mov ecx, [esp + nb010_nn1] ;# esi already loaded with n inc esi sub ecx, esi jecxz .nb010_outerend ;# not last, iterate outer loop once more! mov [esp + nb010_n], esi jmp .nb010_outer.nb010_outerend: ;# check if more outer neighborlists remain mov ecx, [esp + nb010_nri] ;# esi already loaded with n above sub ecx, esi jecxz .nb010_end ;# non-zero, do one more workunit jmp .nb010_threadloop.nb010_end: emms mov eax, [esp + nb010_nouter] mov ebx, [esp + nb010_ninner] mov ecx, [ebp + nb010_outeriter] mov edx, [ebp + nb010_inneriter] mov [ecx], eax mov [edx], ebx mov eax, [esp + nb010_salign] add esp, eax ;# account for stack alignment add esp, 320 pop edi pop esi pop edx pop ecx pop ebx pop eax leave ret .globl nb_kernel010nf_ia32_sse.globl _nb_kernel010nf_ia32_ssenb_kernel010nf_ia32_sse: _nb_kernel010nf_ia32_sse: .equiv nb010nf_p_nri, 8.equiv nb010nf_iinr, 12.equiv nb010nf_jindex, 16.equiv nb010nf_jjnr, 20.equiv nb010nf_shift, 24.equiv nb010nf_shiftvec, 28.equiv nb010nf_fshift, 32.equiv nb010nf_gid, 36.equiv nb010nf_pos, 40.equiv nb010nf_faction, 44.equiv nb010nf_charge, 48.equiv nb010nf_p_facel, 52.equiv nb010nf_p_krf, 56.equiv nb010nf_p_crf, 60.equiv nb010nf_Vc, 64.equiv nb010nf_type, 68.equiv nb010nf_p_ntype, 72.equiv nb010nf_vdwparam, 76.equiv nb010nf_Vvdw, 80.equiv nb010nf_p_tabscale, 84.equiv nb010nf_VFtab, 88.equiv nb010nf_invsqrta, 92.equiv nb010nf_dvda, 96.equiv nb010nf_p_gbtabscale, 100.equiv nb010nf_GBtab, 104.equiv nb010nf_p_nthreads, 108.equiv nb010nf_count, 112.equiv nb010nf_mtx, 116.equiv nb010nf_outeriter, 120.equiv nb010nf_inneriter, 124.equiv nb010nf_work, 128 ;# The mutex (last arg) is not used in assembly. ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb010nf_ix, 0.equiv nb010nf_iy, 16.equiv nb010nf_iz, 32.equiv nb010nf_two, 48.equiv nb010nf_c6, 64.equiv nb010nf_c12, 80.equiv nb010nf_Vvdwtot, 96.equiv nb010nf_half, 112.equiv nb010nf_three, 128.equiv nb010nf_is3, 144.equiv nb010nf_ii3, 148.equiv nb010nf_ntia, 152.equiv nb010nf_innerjjnr, 156.equiv nb010nf_innerk, 160.equiv nb010nf_n, 164.equiv nb010nf_nn1, 168.equiv nb010nf_nri, 172.equiv nb010nf_ntype, 176.equiv nb010nf_nouter, 180.equiv nb010nf_ninner, 184.equiv nb010nf_salign, 188 push ebp mov ebp,esp push eax push ebx push ecx push edx push esi push edi sub esp, 192 ;# local stack space mov eax, esp and eax, 0xf sub esp, eax mov [esp + nb010nf_salign], eax emms ;# Move args passed by reference to stack mov ecx, [ebp + nb010nf_p_nri] mov edi, [ebp + nb010nf_p_ntype] mov ecx, [ecx] mov edi, [edi] mov [esp + nb010nf_nri], ecx mov [esp + nb010nf_ntype], edi ;# zero iteration counters mov eax, 0 mov [esp + nb010nf_nouter], eax mov [esp + nb010nf_ninner], eax⌨️ 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?