nb_kernel331_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,277 行 · 第 1/5 页
S
2,277 行
;# rsqO in xmm7 ;# move ixH1-izH1 to xmm4-xmm6 movaps xmm4, [esp + nb331_ixH1] movaps xmm5, [esp + nb331_iyH1] movaps xmm6, [esp + nb331_izH1] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb331_dxH1], xmm4 movaps [esp + nb331_dyH1], xmm5 movaps [esp + nb331_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 + nb331_ixH2] movaps xmm4, [esp + nb331_iyH2] movaps xmm5, [esp + nb331_izH2] ;# calc dr subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 ;# store dr movaps [esp + nb331_dxH2], xmm3 movaps [esp + nb331_dyH2], xmm4 movaps [esp + nb331_dzH2], xmm5 ;# square it mulps xmm3,xmm3 mulps xmm4,xmm4 mulps xmm5,xmm5 addps xmm5, xmm4 addps xmm5, xmm3 ;# rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 ;# start with rsqO - seed to xmm2 rsqrtps xmm2, xmm7 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm4, [esp + nb331_three] mulps xmm2, xmm7 ;# rsq*lu*lu subps xmm4, xmm2 ;# constant 30-rsq*lu*lu mulps xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm4, [esp + nb331_half] movaps [esp + nb331_rinvO], xmm4 ;# rinvO in xmm4 mulps xmm7, xmm4 movaps [esp + nb331_rO], xmm7 ;# rsqH1 - seed in xmm2 rsqrtps xmm2, xmm6 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm4, [esp + nb331_three] mulps xmm2, xmm6 ;# rsq*lu*lu subps xmm4, xmm2 ;# constant 30-rsq*lu*lu mulps xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm4, [esp + nb331_half] movaps [esp + nb331_rinvH1], xmm4 ;# rinvH1 in xmm4 mulps xmm6, xmm4 movaps [esp + nb331_rH1], xmm6 ;# rsqH2 - seed to xmm2 rsqrtps xmm2, xmm5 movaps xmm3, xmm2 mulps xmm2, xmm2 movaps xmm4, [esp + nb331_three] mulps xmm2, xmm5 ;# rsq*lu*lu subps xmm4, xmm2 ;# constant 30-rsq*lu*lu mulps xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulps xmm4, [esp + nb331_half] movaps [esp + nb331_rinvH2], xmm4 ;# rinvH2 in xmm4 mulps xmm5, xmm4 movaps [esp + nb331_rH2], xmm5 ;# do O interactions ;# rO is still in xmm7 mulps xmm7, [esp + nb331_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 mm0, eax movd mm1, ebx movd mm2, ecx movd mm3, edx mov esi, [ebp + nb331_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] 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 + nb331_two] ;# two*Heps2 movaps xmm0, [esp + nb331_qqO] 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 xmm0, xmm7 ;# fijC=FF*qq ;# at this point mm5 contains vcoul and xmm0 fijC ;# increment vcoul - then we can get rid of mm5 addps xmm5, [esp + nb331_vctot] movaps [esp + nb331_vctot], xmm5 ;# dispersion 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 dispersion 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 dispersion table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# dispersion 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 + nb331_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 + nb331_c6] mulps xmm7, xmm4 ;# fijD mulps xmm5, xmm4 ;# Vvdw6 addps xmm0, xmm7 ;# add to fscal ;# Update Vvdwtot directly addps xmm5, [esp + nb331_Vvdwtot] movaps [esp + nb331_Vvdwtot], xmm5 ;# repulsion 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 repulsion 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 repulsion table movaps xmm6, xmm7 shufps xmm6, xmm3, 136 ;# constant 10001000 shufps xmm7, xmm3, 221 ;# constant 11011101 ;# repulsion 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 + nb331_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 + nb331_c12] mulps xmm7, xmm4 ;# fijD mulps xmm5, xmm4 ;# Vvdw12 addps xmm7, xmm0 ;# add to fscal addps xmm5, [esp + nb331_Vvdwtot] ;# total nonbonded potential in xmm5 xorps xmm4, xmm4 mulps xmm7, [esp + nb331_rinvO] ;# total fscal now in xmm7 mulps xmm7, [esp + nb331_tsc] movaps [esp + nb331_Vvdwtot], xmm5 subps xmm4, xmm7 movaps xmm0, [esp + nb331_dxO] movaps xmm1, [esp + nb331_dyO] movaps xmm2, [esp + nb331_dzO] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# tx in xmm0-xmm2 ;# update O forces movaps xmm3, [esp + nb331_fixO] movaps xmm4, [esp + nb331_fiyO] movaps xmm7, [esp + nb331_fizO] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm7, xmm2 movaps [esp + nb331_fixO], xmm3 movaps [esp + nb331_fiyO], xmm4 movaps [esp + nb331_fizO], xmm7 ;# update j forces with water O movaps [esp + nb331_fjx], xmm0 movaps [esp + nb331_fjy], xmm1 movaps [esp + nb331_fjz], xmm2 ;# Done with O interactions - now H1! movaps xmm7, [esp + nb331_rH1] mulps xmm7, [esp + nb331_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 + nb331_two] ;# two*Heps2 movaps xmm0, [esp + nb331_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 + nb331_vctot] mulps xmm7, [esp + nb331_rinvH1] movaps [esp + nb331_vctot], xmm5 mulps xmm7, [esp + nb331_tsc] subps xmm4, xmm7 movaps xmm0, [esp + nb331_dxH1] movaps xmm1, [esp + nb331_dyH1] movaps xmm2, [esp + nb331_dzH1] mulps xmm0, xmm4 mulps xmm1, xmm4 mulps xmm2, xmm4 ;# update H1 forces movaps xmm3, [esp + nb331_fixH1] movaps xmm4, [esp + nb331_fiyH1] movaps xmm7, [esp + nb331_fizH1] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm7, xmm2 movaps [esp + nb331_fixH1], xmm3 movaps [esp + nb331_fiyH1], xmm4 movaps [esp + nb331_fizH1], xmm7 ;# update j forces with water H1 addps xmm0, [esp + nb331_fjx] addps xmm1, [esp + nb331_fjy] addps xmm2, [esp + nb331_fjz] movaps [esp + nb331_fjx], xmm0 movaps [esp + nb331_fjy], xmm1 movaps [esp + nb331_fjz], xmm2 ;# Done with H1, finally we do H2 interactions movaps xmm7, [esp + nb331_rH2] mulps xmm7, [esp + nb331_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 + nb331_two] ;# two*Heps2 movaps xmm0, [esp + nb331_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 + nb331_vctot] mulps xmm7, [esp + nb331_rinvH2] movaps [esp + nb331_vctot], xmm5 mulps xmm7, [esp + nb331_tsc] subps xmm4, xmm7 movaps xmm0, [esp + nb331_dxH2] movaps xmm1, [esp + nb331_dyH2] movaps xmm2, [esp + nb331_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 + nb331_fixH2] movaps xmm4, [esp + nb331_fiyH2] movaps xmm7, [esp + nb331_fizH2] addps xmm3, xmm0 addps xmm4, xmm1 addps xmm7, xmm2 movaps [esp + nb331_fixH2], xmm3 movaps [esp + nb331_fiyH2], xmm4 movaps [esp + nb331_fizH2], xmm7 mov edi, [ebp + nb331_faction] ;# update j forces addps xmm0, [esp + nb331_fjx] addps xmm1, [esp + nb331_fjy] addps xmm2, [esp + nb331_fjz] movlps xmm4, [edi + eax*4] movlps xmm7, [edi + ecx*4] movhps xmm4, [edi + ebx*4] movhps xmm7, [edi + edx*4] movaps xmm3, xmm4 shufps xmm3, xmm7, 136 ;# constant 10001000 shufps xmm4, xmm7, 221 ;# constant 11011101 ;# xmm3 has fjx, xmm4 has fjy ⌨️ 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?