nb_kernel332_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,165 行 · 第 1/5 页
S
2,165 行
;# 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 + nb332_ixO] movaps xmm1, [esp + nb332_iyO] movaps xmm2, [esp + nb332_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 + nb332_jxO], xmm3 movaps [esp + nb332_jyO], xmm4 movaps [esp + nb332_jzO], xmm5 subps xmm0, xmm3 subps xmm1, xmm4 subps xmm2, xmm5 movaps [esp + nb332_dxOO], xmm0 movaps [esp + nb332_dyOO], xmm1 movaps [esp + nb332_dzOO], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 ;# have rsq in xmm0 ;# do invsqrt rsqrtps xmm1, xmm0 movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [esp + nb332_three] mulps xmm1, xmm0 subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [esp + nb332_half] ;# rinv iO - j water movaps xmm1, xmm3 mulps xmm1, xmm0 ;# xmm1=r movaps xmm0, xmm3 ;# xmm0=rinv mulps xmm1, [esp + nb332_tsc] movhlps xmm2, xmm1 cvttps2pi mm6, xmm1 cvttps2pi mm7, xmm2 ;# mm6/mm7 contain lu indices cvtpi2ps xmm3, mm6 cvtpi2ps xmm2, mm7 movlhps xmm3, xmm2 subps xmm1, xmm3 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd ebx, mm6 movd ecx, mm7 psrlq mm7, 32 movd edx, mm7 ;# table indices in ebx,ecx,edx mov esi, [ebp + nb332_VFtab] lea ebx, [ebx + ebx*2] lea ecx, [ecx + ecx*2] lea edx, [edx + edx*2] movlps xmm5, [esi + ebx*4] movlps xmm7, [esi + ecx*4] movhps xmm7, [esi + edx*4] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [esi + ebx*4 + 8] movlps xmm3, [esi + ecx*4 + 8] movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# coulomb table ready, in xmm4-xmm7 mulps xmm6, xmm1 ;# xmm6=Geps mulps xmm7, xmm2 ;# xmm7=Heps2 addps xmm5, xmm6 addps xmm5, xmm7 ;# xmm5=Fp mulps xmm7, [esp + nb332_two] ;# two*Heps2 xorps xmm3, xmm3 ;# fetch charges to xmm3 (temporary) movss xmm3, [esp + nb332_qqOO] movhps xmm3, [esp + nb332_qqOH] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV mulps xmm3, xmm7 ;# fijC=FF*qq ;# at this point xmm5 contains vcoul and xmm3 fijC addps xmm5, [esp + nb332_vctot] movaps [esp + nb332_vctot], xmm5 ;# put scalar force on stack temporarily movaps [esp + nb332_fstmp], xmm3 ;# dispersion movss xmm4, [esi + ebx*4 + 16] movss xmm5, [esi + ebx*4 + 20] movss xmm6, [esi + ebx*4 + 24] movss xmm7, [esi + ebx*4 + 28] ;# 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 + nb332_two] ;# two*Heps2 addss xmm7, xmm6 addss xmm7, xmm5 ;# xmm7=FF mulss xmm5, xmm1 ;# xmm5=eps*Fp addss xmm5, xmm4 ;# xmm5=VV xorps xmm4, xmm4 movss xmm4, [esp + nb332_c6] mulps xmm7, xmm4 ;# fijD mulps xmm5, xmm4 ;# Vvdw6 addps xmm7, [esp + nb332_fstmp] ;# add to fscal ;# put scalar force on stack Update Vvdwtot directly addps xmm5, [esp + nb332_Vvdwtot] movaps [esp + nb332_fstmp], xmm7 movaps [esp + nb332_Vvdwtot], xmm5 ;# repulsion movss xmm4, [esi + ebx*4 + 32] movss xmm5, [esi + ebx*4 + 36] movss xmm6, [esi + ebx*4 + 40] movss xmm7, [esi + ebx*4 + 44] ;# 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 + nb332_two] ;# two*Heps2 addss xmm7, xmm6 addss xmm7, xmm5 ;# xmm7=FF mulss xmm5, xmm1 ;# xmm5=eps*Fp addss xmm5, xmm4 ;# xmm5=VV xorps xmm4, xmm4 movss xmm4, [esp + nb332_c12] mulps xmm7, xmm4 ;# fijR mulps xmm5, xmm4 ;# Vvdw12 addps xmm7, [esp + nb332_fstmp] addps xmm5, [esp + nb332_Vvdwtot] movaps [esp + nb332_Vvdwtot], xmm5 xorps xmm1, xmm1 mulps xmm7, [esp + nb332_tsc] mulps xmm7, xmm0 subps xmm1, xmm7 movaps xmm0, xmm1 movaps xmm2, xmm1 mulps xmm0, [esp + nb332_dxOO] mulps xmm1, [esp + nb332_dyOO] mulps xmm2, [esp + nb332_dzOO] ;# 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 + nb332_fjxO], xmm3 movaps [esp + nb332_fjyO], xmm4 movaps [esp + nb332_fjzO], xmm5 addps xmm0, [esp + nb332_fixO] addps xmm1, [esp + nb332_fiyO] addps xmm2, [esp + nb332_fizO] movaps [esp + nb332_fixO], xmm0 movaps [esp + nb332_fiyO], xmm1 movaps [esp + nb332_fizO], xmm2 ;# done with i O Now do i H1 & H2 simultaneously first get i particle coords: movaps xmm0, [esp + nb332_ixH1] movaps xmm1, [esp + nb332_iyH1] movaps xmm2, [esp + nb332_izH1] movaps xmm3, [esp + nb332_ixH2] movaps xmm4, [esp + nb332_iyH2] movaps xmm5, [esp + nb332_izH2] subps xmm0, [esp + nb332_jxO] subps xmm1, [esp + nb332_jyO] subps xmm2, [esp + nb332_jzO] subps xmm3, [esp + nb332_jxO] subps xmm4, [esp + nb332_jyO] subps xmm5, [esp + nb332_jzO] movaps [esp + nb332_dxH1O], xmm0 movaps [esp + nb332_dyH1O], xmm1 movaps [esp + nb332_dzH1O], xmm2 movaps [esp + nb332_dxH2O], xmm3 movaps [esp + nb332_dyH2O], xmm4 movaps [esp + nb332_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 ;# start with H1, save H2 data movaps [esp + nb332_rsqH2O], xmm4 ;# do invsqrt rsqrtps xmm1, xmm0 rsqrtps xmm5, xmm4 movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb332_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 + nb332_half] ;# rinv H1 - j water mulps xmm7, [esp + nb332_half] ;# rinv H2 - j water ;# start with H1, save H2 data movaps [esp + nb332_rinvH2O], xmm7 movaps xmm1, xmm3 mulps xmm1, xmm0 ;# xmm1=r movaps xmm0, xmm3 ;# xmm0=rinv mulps xmm1, [esp + nb332_tsc] movhlps xmm2, xmm1 cvttps2pi mm6, xmm1 cvttps2pi mm7, xmm2 ;# mm6/mm7 contain lu indices cvtpi2ps xmm3, mm6 cvtpi2ps xmm2, mm7 movlhps xmm3, xmm2 subps xmm1, xmm3 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd ebx, mm6 movd ecx, mm7 psrlq mm7, 32 movd edx, mm7 ;# table indices in ebx,ecx,edx lea ebx, [ebx + ebx*2] lea ecx, [ecx + ecx*2] lea edx, [edx + edx*2] movlps xmm5, [esi + ebx*4] movlps xmm7, [esi + ecx*4] movhps xmm7, [esi + edx*4] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [esi + ebx*4 + 8] movlps xmm3, [esi + ecx*4 + 8] movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# coulomb table ready, in xmm4-xmm7 mulps xmm6, xmm1 ;# xmm6=Geps mulps xmm7, xmm2 ;# xmm7=Heps2 addps xmm5, xmm6 addps xmm5, xmm7 ;# xmm5=Fp mulps xmm7, [esp + nb332_two] ;# two*Heps2 xorps xmm3, xmm3 ;# fetch charges to xmm3 (temporary) movss xmm3, [esp + nb332_qqOH] movhps xmm3, [esp + nb332_qqHH] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV mulps xmm3, xmm7 ;# fijC=FF*qq ;# at this point xmm5 contains vcoul and xmm3 fijC addps xmm5, [esp + nb332_vctot] movaps [esp + nb332_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb332_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 mulps xmm0, [esp + nb332_dxH1O] mulps xmm1, [esp + nb332_dyH1O] mulps xmm2, [esp + nb332_dzH1O] ;# update forces H1 - j water movaps xmm3, [esp + nb332_fjxO] movaps xmm4, [esp + nb332_fjyO] movaps xmm5, [esp + nb332_fjzO] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 movaps [esp + nb332_fjxO], xmm3 movaps [esp + nb332_fjyO], xmm4 movaps [esp + nb332_fjzO], xmm5 addps xmm0, [esp + nb332_fixH1] addps xmm1, [esp + nb332_fiyH1] addps xmm2, [esp + nb332_fizH1] movaps [esp + nb332_fixH1], xmm0 movaps [esp + nb332_fiyH1], xmm1 movaps [esp + nb332_fizH1], xmm2 ;# do table for H2 - j water interaction movaps xmm0, [esp + nb332_rinvH2O] movaps xmm1, [esp + nb332_rsqH2O] mulps xmm1, xmm0 ;# xmm0=rinv, xmm1=r mulps xmm1, [esp + nb332_tsc] movhlps xmm2, xmm1 cvttps2pi mm6, xmm1 cvttps2pi mm7, xmm2 ;# mm6/mm7 contain lu indices cvtpi2ps xmm3, mm6 cvtpi2ps xmm2, mm7 movlhps xmm3, xmm2 subps xmm1, xmm3 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd ebx, mm6 movd ecx, mm7 psrlq mm7, 32 movd edx, mm7 ;# table indices in ebx,ecx,edx lea ebx, [ebx + ebx*2] lea ecx, [ecx + ecx*2] lea edx, [edx + edx*2] movlps xmm5, [esi + ebx*4] movlps xmm7, [esi + ecx*4] movhps xmm7, [esi + edx*4] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [esi + ebx*4 + 8] movlps xmm3, [esi + ecx*4 + 8] movhps xmm3, [esi + edx*4 + 8] ;# other half of coulomb table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# coulomb table ready, in xmm4-xmm7 mulps xmm6, xmm1 ;# xmm6=Geps mulps xmm7, xmm2 ;# xmm7=Heps2 addps xmm5, xmm6 addps xmm5, xmm7 ;# xmm5=Fp mulps xmm7, [esp + nb332_two] ;# two*Heps2 xorps xmm3, xmm3 ;# fetch charges to xmm3 (temporary) movss xmm3, [esp + nb332_qqOH] movhps xmm3, [esp + nb332_qqHH] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm3 ;# vcoul=qq*VV mulps xmm3, xmm7 ;# fijC=FF*qq ;# at this point xmm5 contains vcoul and xmm3 fijC addps xmm5, [esp + nb332_vctot] movaps [esp + nb332_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb332_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 mulps xmm0, [esp + nb332_dxH2O] mulps xmm1, [esp + nb332_dyH2O] mulps xmm2, [esp + nb332_dzH2O] movaps xmm3, [esp + nb332_fjxO] movaps xmm4, [esp + nb332_fjyO] movaps xmm5, [esp + nb332_fjzO] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 mov esi, [ebp + nb332_faction] movaps [esp + nb332_fjxO], xmm3 movaps [esp + nb332_fjyO], xmm4 movaps [esp + nb332_fjzO], xmm5 addps xmm0, [esp + nb332_fixH2] addps xmm1, [esp + nb332_fiyH2] addps xmm2, [esp + nb332_fizH2] movaps [esp + nb332_fixH2], xmm0 movaps [esp + nb332_fiyH2], xmm1 movaps [esp + nb332_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 + nb332_⌨️ 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?