nb_kernel232_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,146 行 · 第 1/5 页
S
2,146 行
jl .nb232_single_check jmp .nb232_unroll_loop.nb232_single_check: add dword ptr [esp + nb232_innerk], 4 jnz .nb232_single_loop jmp .nb232_updateouterdata.nb232_single_loop: mov edx, [esp + nb232_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] add dword ptr [esp + nb232_innerjjnr], 4 mov esi, [ebp + nb232_pos] lea eax, [eax + eax*2] ;# fetch j coordinates xorps xmm3, xmm3 xorps xmm4, xmm4 xorps xmm5, xmm5 movss xmm3, [esi + eax*4] ;# jxO - - - movss xmm4, [esi + eax*4 + 4] ;# jyO - - - movss xmm5, [esi + eax*4 + 8] ;# jzO - - - movlps xmm6, [esi + eax*4 + 12] ;# xmm6 = jxH1 jyH1 - - movss xmm7, [esi + eax*4 + 20] ;# xmm7 = jzH1 - - - movhps xmm6, [esi + eax*4 + 24] ;# xmm6 = jxH1 jyH1 jxH2 jyH2 movss xmm2, [esi + eax*4 + 32] ;# xmm2 = jzH2 - - - ;# have all coords, time for some shuffling. shufps xmm6, xmm6, 216 ;# constant 11011000 ;# xmm6 = jxH1 jxH2 jyH1 jyH2 unpcklps xmm7, xmm2 ;# xmm7 = jzH1 jzH2 - - movaps xmm0, [esp + nb232_ixO] movaps xmm1, [esp + nb232_iyO] movaps xmm2, [esp + nb232_izO] movlhps xmm3, xmm6 ;# xmm3 = jxO 0 jxH1 jxH2 shufps xmm4, xmm6, 228 ;# constant 11100100 ;# xmm4 = jyO 0 jyH1 jyH2 shufps xmm5, xmm7, 68 ;# constant 01000100 ;# xmm5 = jzO 0 jzH1 jzH2 ;# store all j coordinates in jO movaps [esp + nb232_jxO], xmm3 movaps [esp + nb232_jyO], xmm4 movaps [esp + nb232_jzO], xmm5 subps xmm0, xmm3 subps xmm1, xmm4 subps xmm2, xmm5 movaps [esp + nb232_dxOO], xmm0 movaps [esp + nb232_dyOO], xmm1 movaps [esp + nb232_dzOO], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 ;# have rsq in xmm0 movaps [esp + nb232_rsqOO], xmm0 movaps xmm6, xmm0 ;# do invsqrt rsqrtps xmm1, xmm0 mulps xmm6, [esp + nb232_krf] ;# xmm6=krsq movaps xmm2, xmm1 movaps xmm7, xmm6 mulps xmm1, xmm1 movaps xmm3, [esp + nb232_three] mulps xmm1, xmm0 subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [esp + nb232_half] ;# rinv iO - j water in xmm3 movaps [esp + nb232_rinvOO], xmm3 addps xmm6, xmm3 ;# xmm6=rinv+ krsq mulps xmm7, [esp + nb232_two] subps xmm6, [esp + nb232_crf] ;# xmm6=rinv+ krsq-crf movaps xmm0, xmm3 ;# rinv subps xmm3, xmm7 ;# xmm3=rinv-2*krsq xorps xmm4, xmm4 ;# fetch charges to xmm4 (temporary) movss xmm4, [esp + nb232_qqOO] movhps xmm4, [esp + nb232_qqOH] mulps xmm6, xmm4 ;# vcoul mulps xmm3, xmm4 ;# coul part of fs mulps xmm3, xmm0 movaps [esp + nb232_fstmp], xmm3 ;# save it addps xmm6, [esp + nb232_vctot] movaps [esp + nb232_vctot], xmm6 movaps xmm0, [esp + nb232_rinvOO] movss xmm1, xmm0 mulss xmm1, [esp + nb232_rsqOO] ;# xmm1=r mulss xmm1, [esp + nb232_tsc] cvttps2pi mm6, xmm1 cvtpi2ps xmm3, mm6 subss xmm1, xmm3 ;# xmm1=eps movss xmm2, xmm1 mulss xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 3 movd mm0, eax mov esi, [ebp + nb232_VFtab] movd eax, mm6 ;# dispersion movlps xmm5, [esi + eax*4] movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [esi + eax*4 + 8] movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# dispersion table ready, in xmm4-xmm7 mulss xmm6, xmm1 ;# xmm6=Geps mulss xmm7, xmm2 ;# xmm7=Heps2 addss xmm5, xmm6 addss xmm5, xmm7 ;# xmm5=Fp mulss xmm7, [esp + nb232_two] ;# two*Heps2 addss xmm7, xmm6 addss xmm7, xmm5 ;# xmm7=FF mulss xmm5, xmm1 ;# xmm5=eps*Fp addss xmm5, xmm4 ;# xmm5=VV movss xmm4, [esp + nb232_c6] mulss xmm7, xmm4 ;# fijD mulss xmm5, xmm4 ;# Vvdw6 movss xmm3, [esp + nb232_fstmp] mulps xmm7, [esp + nb232_tsc] subss xmm3, xmm7 movss [esp + nb232_fstmp], xmm3 addss xmm5, [esp + nb232_Vvdwtot] movss [esp + nb232_Vvdwtot], xmm5 ;# repulsion movlps xmm5, [esi + eax*4 + 16] movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [esi + eax*4 + 24] movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# table ready, in xmm4-xmm7 mulss xmm6, xmm1 ;# xmm6=Geps mulss xmm7, xmm2 ;# xmm7=Heps2 addss xmm5, xmm6 addss xmm5, xmm7 ;# xmm5=Fp mulss xmm7, [esp + nb232_two] ;# two*Heps2 addss xmm7, xmm6 addss xmm7, xmm5 ;# xmm7=FF mulss xmm5, xmm1 ;# xmm5=eps*Fp addss xmm5, xmm4 ;# xmm5=VV movss xmm4, [esp + nb232_c12] mulss xmm7, xmm4 ;# fijR mulss xmm5, xmm4 ;# Vvdw12 movaps xmm3, [esp + nb232_fstmp] mulss xmm7, [esp + nb232_tsc] subss xmm3, xmm7 addss xmm5, [esp + nb232_Vvdwtot] movss [esp + nb232_Vvdwtot], xmm5 mulps xmm0, xmm3 movaps xmm1, xmm0 movaps xmm2, xmm0 mulps xmm0, [esp + nb232_dxOO] mulps xmm1, [esp + nb232_dyOO] mulps xmm2, [esp + nb232_dzOO] movd eax, mm0 ;# initial update for j forces xorps xmm3, xmm3 xorps xmm4, xmm4 xorps xmm5, xmm5 subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 movaps [esp + nb232_fjxO], xmm3 movaps [esp + nb232_fjyO], xmm4 movaps [esp + nb232_fjzO], xmm5 addps xmm0, [esp + nb232_fixO] addps xmm1, [esp + nb232_fiyO] addps xmm2, [esp + nb232_fizO] movaps [esp + nb232_fixO], xmm0 movaps [esp + nb232_fiyO], xmm1 movaps [esp + nb232_fizO], xmm2 ;# done with i O Now do i H1 & H2 simultaneously first get i particle coords: movaps xmm0, [esp + nb232_ixH1] movaps xmm1, [esp + nb232_iyH1] movaps xmm2, [esp + nb232_izH1] movaps xmm3, [esp + nb232_ixH2] movaps xmm4, [esp + nb232_iyH2] movaps xmm5, [esp + nb232_izH2] subps xmm0, [esp + nb232_jxO] subps xmm1, [esp + nb232_jyO] subps xmm2, [esp + nb232_jzO] subps xmm3, [esp + nb232_jxO] subps xmm4, [esp + nb232_jyO] subps xmm5, [esp + nb232_jzO] movaps [esp + nb232_dxH1O], xmm0 movaps [esp + nb232_dyH1O], xmm1 movaps [esp + nb232_dzH1O], xmm2 movaps [esp + nb232_dxH2O], xmm3 movaps [esp + nb232_dyH2O], xmm4 movaps [esp + nb232_dzH2O], xmm5 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm4, xmm3 addps xmm0, xmm2 ;# have rsqH1 in xmm0 addps xmm4, xmm5 ;# have rsqH2 in xmm4 ;# do invsqrt rsqrtps xmm1, xmm0 rsqrtps xmm5, xmm4 movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb232_three] movaps xmm7, xmm3 mulps xmm1, xmm0 mulps xmm5, xmm4 subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb232_half] ;# rinv H1 - j water mulps xmm7, [esp + nb232_half] ;# rinv H2 - j water mulps xmm0, [esp + nb232_krf] ;# krsq mulps xmm4, [esp + nb232_krf] ;# krsq ;# assemble charges in xmm6 xorps xmm6, xmm6 movss xmm6, [esp + nb232_qqOH] movhps xmm6, [esp + nb232_qqHH] movaps xmm1, xmm0 movaps xmm5, xmm4 addps xmm0, xmm3 ;# krsq+ rinv addps xmm4, xmm7 ;# krsq+ rinv subps xmm0, [esp + nb232_crf] subps xmm4, [esp + nb232_crf] mulps xmm1, [esp + nb232_two] mulps xmm5, [esp + nb232_two] mulps xmm0, xmm6 ;# vcoul mulps xmm4, xmm6 ;# vcoul addps xmm4, xmm0 addps xmm4, [esp + nb232_vctot] movaps [esp + nb232_vctot], xmm4 movaps xmm0, xmm3 movaps xmm4, xmm7 mulps xmm3, xmm3 mulps xmm7, xmm7 subps xmm0, xmm1 subps xmm4, xmm5 mulps xmm0, xmm6 mulps xmm4, xmm6 mulps xmm0, xmm3 ;# fscal mulps xmm7, xmm4 ;# fscal movaps xmm1, xmm0 movaps xmm2, xmm0 mulps xmm0, [esp + nb232_dxH1O] mulps xmm1, [esp + nb232_dyH1O] mulps xmm2, [esp + nb232_dzH1O] ;# update forces H1 - j water movaps xmm3, [esp + nb232_fjxO] movaps xmm4, [esp + nb232_fjyO] movaps xmm5, [esp + nb232_fjzO] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 movaps [esp + nb232_fjxO], xmm3 movaps [esp + nb232_fjyO], xmm4 movaps [esp + nb232_fjzO], xmm5 addps xmm0, [esp + nb232_fixH1] addps xmm1, [esp + nb232_fiyH1] addps xmm2, [esp + nb232_fizH1] movaps [esp + nb232_fixH1], xmm0 movaps [esp + nb232_fiyH1], xmm1 movaps [esp + nb232_fizH1], xmm2 ;# do forces H2 - j water movaps xmm0, xmm7 movaps xmm1, xmm7 movaps xmm2, xmm7 mulps xmm0, [esp + nb232_dxH2O] mulps xmm1, [esp + nb232_dyH2O] mulps xmm2, [esp + nb232_dzH2O] movaps xmm3, [esp + nb232_fjxO] movaps xmm4, [esp + nb232_fjyO] movaps xmm5, [esp + nb232_fjzO] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 mov esi, [ebp + nb232_faction] movaps [esp + nb232_fjxO], xmm3 movaps [esp + nb232_fjyO], xmm4 movaps [esp + nb232_fjzO], xmm5 addps xmm0, [esp + nb232_fixH2] addps xmm1, [esp + nb232_fiyH2] addps xmm2, [esp + nb232_fizH2] movaps [esp + nb232_fixH2], xmm0 movaps [esp + nb232_fiyH2], xmm1 movaps [esp + nb232_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 + nb232_fjxO] movaps xmm4, [esp + nb232_fjyO] movaps xmm5, [esp + nb232_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 + nb232_innerk] jz .nb232_updateouterdata jmp .nb232_single_loop.nb232_updateouterdata: mov ecx, [esp + nb232_ii3] mov edi, [ebp + nb232_faction] mov esi, [ebp + nb232_fshift] mov edx, [esp + nb232_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movaps xmm0, [esp + nb232_fixO] movaps xmm1, [esp + nb232_fiyO] movaps xmm2, [esp + nb232_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 + nb232_fixH1] movaps xmm1, [esp + nb232_fiyH1] movaps xmm2, [esp + nb232_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⌨️ 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?