nb_kernel333_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,386 行 · 第 1/5 页
S
2,386 行
;# move ixH1-izH1 to xmm4-xmm6 movaps xmm4, [esp + nb333_ixH1] movaps xmm5, [esp + nb333_iyH1] movaps xmm6, [esp + nb333_izH1] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb333_dxH1], xmm4 movaps [esp + nb333_dyH1], xmm5 movaps [esp + nb333_dzH1], xmm6 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm6, xmm5 addps xmm6, xmm4 ;# rsqH1 in xmm6 ;# move ixH2-izH2 to xmm3-xmm5 movaps xmm3, [esp + nb333_ixH2] movaps xmm4, [esp + nb333_iyH2] movaps xmm5, [esp + nb333_izH2] ;# calc dr subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 ;# store dr movaps [esp + nb333_dxH2], xmm3 movaps [esp + nb333_dyH2], xmm4 movaps [esp + nb333_dzH2], xmm5 ;# square it mulps xmm3,xmm3 mulps xmm4,xmm4 mulps xmm5,xmm5 addps xmm5, xmm4 addps xmm5, xmm3 ;# move ixM-izM to xmm2-xmm4 movaps xmm3, [esp + nb333_iyM] movaps xmm4, [esp + nb333_izM] subps xmm3, xmm1 subps xmm4, xmm2 movaps xmm2, [esp + nb333_ixM] subps xmm2, xmm0 ;# store dr movaps [esp + nb333_dxM], xmm2 movaps [esp + nb333_dyM], xmm3 movaps [esp + nb333_dzM], xmm4 ;# square it mulps xmm2,xmm2 mulps xmm3,xmm3 mulps xmm4,xmm4 addps xmm4, xmm3 addps xmm4, xmm2 ;# rsqM in xmm4, rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 ;# rsqH1 - seed in xmm2 rsqrtps xmm2, xmm6 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm0, [esp + nb333_three] mulps xmm2, xmm6 ;# rsq*lu*lu subps xmm0, xmm2 ;# constant 30-rsq*lu*lu mulps xmm0, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm0, [esp + nb333_half] movaps [esp + nb333_rinvH1], xmm0 ;# rinvH1 mulps xmm6, xmm0 movaps [esp + nb333_rH1], xmm6 ;# rsqH2 - seed to xmm2 rsqrtps xmm2, xmm5 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm0, [esp + nb333_three] mulps xmm2, xmm5 ;# rsq*lu*lu subps xmm0, xmm2 ;# constant 30-rsq*lu*lu mulps xmm0, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm0, [esp + nb333_half] movaps [esp + nb333_rinvH2], xmm0 ;# rinvH2 mulps xmm5, xmm0 movaps [esp + nb333_rH2], xmm5 ;# rsqM - seed to xmm2 rsqrtps xmm2, xmm4 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm0, [esp + nb333_three] mulps xmm2, xmm4 ;# rsq*lu*lu subps xmm0, xmm2 ;# constant 30-rsq*lu*lu mulps xmm0, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm0, [esp + nb333_half] movaps [esp + nb333_rinvM], xmm0 ;# rinvM mulps xmm4, xmm0 movaps [esp + nb333_rM], xmm4 ;# Do the O LJ table interaction directly. rsqrtps xmm2, xmm7 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm0, [esp + nb333_three] mulps xmm2, xmm7 ;# rsq*lu*lu subps xmm0, xmm2 ;# constant 30-rsq*lu*lu mulps xmm0, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm0, [esp + nb333_half] ;# rinv movaps xmm1, xmm0 mulps xmm1, xmm7 ;# xmm1=r mulps xmm1, [esp + nb333_tsc] ;# r*tabscale 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 mm0, eax movd mm1, ebx movd mm2, ecx movd mm3, edx mov esi, [ebp + nb333_VFtab] movd eax, mm6 psrlq mm6, 32 movd ecx, mm7 psrlq mm7, 32 movd ebx, mm6 movd edx, mm7 lea eax, [eax + eax*2] lea ebx, [ebx + ebx*2] lea ecx, [ecx + ecx*2] lea edx, [edx + edx*2] ;# load dispersion table data into xmm4-xmm7 movlps xmm5, [esi + eax*4 + 16] movlps xmm7, [esi + ecx*4 + 16] movhps xmm5, [esi + ebx*4 + 16] movhps xmm7, [esi + edx*4 + 16] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [esi + eax*4 + 24] movlps xmm3, [esi + ecx*4 + 24] movhps xmm7, [esi + ebx*4 + 24] movhps xmm3, [esi + edx*4 + 24] ;# other half of coulomb table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# dispersion table YFGH 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 + nb333_two] ;# two*Heps2 addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV movaps xmm4, [esp + nb333_c6] mulps xmm7, xmm4 ;# fijD mulps xmm5, xmm4 ;# Vvdw6 ;# put scalar force on stack (borrow rinvO) ;# Update Vvdwtot directly addps xmm5, [esp + nb333_Vvdwtot] movaps [esp + nb333_rinvO], xmm7 ;# fscal movaps [esp + nb333_Vvdwtot], xmm5 ;# load repulsion table data into xmm4-xmm7 movlps xmm5, [esi + eax*4 + 32] movlps xmm7, [esi + ecx*4 + 32] movhps xmm5, [esi + ebx*4 + 32] movhps xmm7, [esi + edx*4 + 32] ;# got half coulomb table movaps xmm4, xmm5 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm5, xmm7, 221 ;# constant 11011101 movlps xmm7, [esi + eax*4 + 40] movlps xmm3, [esi + ecx*4 + 40] movhps xmm7, [esi + ebx*4 + 40] movhps xmm3, [esi + edx*4 + 40] ;# other half of coulomb table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# repulsion table YFGH 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 + nb333_two] ;# two*Heps2 addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV movaps xmm4, [esp + nb333_c12] mulps xmm7, xmm4 ;# fijR mulps xmm5, xmm4 ;# Vvdw12 addps xmm7, [esp + nb333_rinvO] ;# fscal was temp. stored in rinvO addps xmm5, [esp + nb333_Vvdwtot] movaps [esp + nb333_Vvdwtot], xmm5 xorps xmm1, xmm1 mulps xmm7, [esp + nb333_tsc] mulps xmm7, xmm0 subps xmm1, xmm7 ;# fscal movaps xmm3, [esp + nb333_dxO] movaps xmm4, [esp + nb333_dyO] movaps xmm5, [esp + nb333_dzO] mulps xmm3, xmm1 mulps xmm4, xmm1 mulps xmm5, xmm1 ;# tx in xmm3-xmm5 ;# update O forces movaps xmm0, [esp + nb333_fixO] movaps xmm1, [esp + nb333_fiyO] movaps xmm2, [esp + nb333_fizO] addps xmm0, xmm3 addps xmm1, xmm4 addps xmm2, xmm5 movaps [esp + nb333_fixO], xmm0 movaps [esp + nb333_fiyO], xmm1 movaps [esp + nb333_fizO], xmm2 ;# update j forces with water O movaps [esp + nb333_fjx], xmm3 movaps [esp + nb333_fjy], xmm4 movaps [esp + nb333_fjz], xmm5 ;# Do H1 interaction mov esi, [ebp + nb333_VFtab] movaps xmm7, [esp + nb333_rH1] mulps xmm7, [esp + nb333_tsc] movhlps xmm4, xmm7 cvttps2pi mm6, xmm7 cvttps2pi mm7, xmm4 ;# mm6/mm7 contain lu indices cvtpi2ps xmm3, mm6 cvtpi2ps xmm4, mm7 movlhps xmm3, xmm4 subps xmm7, xmm3 movaps xmm1, xmm7 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd eax, mm6 psrlq mm6, 32 movd ecx, mm7 psrlq mm7, 32 movd ebx, mm6 movd edx, mm7 lea eax, [eax + eax*2] lea ebx, [ebx + ebx*2] lea ecx, [ecx + ecx*2] lea edx, [edx + edx*2] movlps xmm5, [esi + eax*4] movlps xmm7, [esi + ecx*4] movhps xmm5, [esi + ebx*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 + eax*4 + 8] movlps xmm3, [esi + ecx*4 + 8] movhps xmm7, [esi + ebx*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 + nb333_two] ;# two*Heps2 movaps xmm0, [esp + nb333_qqH] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm0 ;# vcoul=qq*VV mulps xmm7, xmm0 ;# fijC=FF*qq ;# at this point mm5 contains vcoul and xmm7 fijC ;# increment vcoul xorps xmm4, xmm4 addps xmm5, [esp + nb333_vctot] mulps xmm7, [esp + nb333_rinvH1] movaps [esp + nb333_vctot], xmm5 mulps xmm7, [esp + nb333_tsc] subps xmm4, xmm7 movaps xmm0, [esp + nb333_dxH1] movaps xmm1, [esp + nb333_dyH1] movaps xmm2, [esp + nb333_dzH1] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# update H1 forces movaps xmm3, [esp + nb333_fixH1] movaps xmm4, [esp + nb333_fiyH1] movaps xmm7, [esp + nb333_fizH1] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm7, xmm2 movaps [esp + nb333_fixH1], xmm3 movaps [esp + nb333_fiyH1], xmm4 movaps [esp + nb333_fizH1], xmm7 ;# update j forces with water H1 addps xmm0, [esp + nb333_fjx] addps xmm1, [esp + nb333_fjy] addps xmm2, [esp + nb333_fjz] movaps [esp + nb333_fjx], xmm0 movaps [esp + nb333_fjy], xmm1 movaps [esp + nb333_fjz], xmm2 ;# Done with H1, do H2 interactions movaps xmm7, [esp + nb333_rH2] mulps xmm7, [esp + nb333_tsc] movhlps xmm4, xmm7 cvttps2pi mm6, xmm7 cvttps2pi mm7, xmm4 ;# mm6/mm7 contain lu indices cvtpi2ps xmm3, mm6 cvtpi2ps xmm4, mm7 movlhps xmm3, xmm4 subps xmm7, xmm3 movaps xmm1, xmm7 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd eax, mm6 psrlq mm6, 32 movd ecx, mm7 psrlq mm7, 32 movd ebx, mm6 movd edx, mm7 lea eax, [eax + eax*2] lea ebx, [ebx + ebx*2] lea ecx, [ecx + ecx*2] lea edx, [edx + edx*2] movlps xmm5, [esi + eax*4] movlps xmm7, [esi + ecx*4] movhps xmm5, [esi + ebx*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 + eax*4 + 8] movlps xmm3, [esi + ecx*4 + 8] movhps xmm7, [esi + ebx*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 + nb333_two] ;# two*Heps2 movaps xmm0, [esp + nb333_qqH] addps xmm7, xmm6 addps xmm7, xmm5 ;# xmm7=FF mulps xmm5, xmm1 ;# xmm5=eps*Fp addps xmm5, xmm4 ;# xmm5=VV mulps xmm5, xmm0 ;# vcoul=qq*VV mulps xmm7, xmm0 ;# fijC=FF*qq ;# at this point mm5 contains vcoul and xmm0 fijC ;# increment vcoul xorps xmm4, xmm4 addps xmm5, [esp + nb333_vctot] mulps xmm7, [esp + nb333_rinvH2] movaps [esp + nb333_vctot], xmm5 mulps xmm7, [esp + nb333_tsc] subps xmm4, xmm7 movaps xmm0, [esp + nb333_dxH2] movaps xmm1, [esp + nb333_dyH2] movaps xmm2, [esp + nb333_dzH2] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 movd eax, mm0 movd ebx, mm1 movd ecx, mm2 movd edx, mm3 ;# update H2 forces movaps xmm3, [esp + nb333_fixH2] movaps xmm4, [esp + nb333_fiyH2] movaps xmm7, [esp + nb333_fizH2] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm7, xmm2 movaps [esp + nb333_fixH2], xmm3 movaps [esp + nb333_fiyH2], xmm4 movaps [esp + nb333_fizH2], xmm7 addps xmm0, [esp + nb333_fjx] addps xmm1, [esp + nb333_fjy] addps xmm2, [esp + nb333_fjz] movaps [esp + nb333_fjx], xmm0 movaps [esp + nb333_fjy], xmm1 movaps [esp + nb333_fjz], xmm2 ;# Done with H2, do M interactions movaps xmm7, [esp + nb333_rM] mulps xmm7, [esp + nb333_tsc] movhlps xmm4, xmm7 cvttps2pi mm6, xmm7 cvttps2pi mm7, xmm4 ;# mm6/mm7 contain lu indices cvtpi2ps xmm3, mm6 cvtpi2ps xmm4, mm7 movlhps xmm3, xmm4 subps xmm7, xmm3 movaps xmm1, xmm7 ;# xmm1=eps movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=eps2 pslld mm6, 2 pslld mm7, 2 movd eax, mm6 psrlq mm6, 32 movd ecx, mm7 psrlq mm7, 32 movd ebx, mm6 movd edx, mm7 lea eax, [eax + eax*2] lea ebx, [ebx + ebx*2] lea ecx, [ecx + ecx*2] lea edx, [edx + edx*2] movlps xmm5, [esi + eax*4] movlps xmm7, [esi + ecx*4] movhps xmm5, [esi + ebx*4] movhps xmm7, [esi + edx*4] ;# got half coulomb table
⌨️ 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?