📄 nb_kernel332_ia32_3dnow.intel_syntax.s
字号:
pfmul mm0,mm1 ;# mm0=r pfmul mm0, [esp + nb332_tsc] pf2iw mm4, mm0 movq [esp + nb332_n1], mm4 pi2fd mm4,mm4 pfsub mm0, mm4 ;# now mm0 is eps and mm4 n0 movq mm2, mm0 pfmul mm2, mm2 ;# mm0 is eps, mm2 eps2 ;# coulomb table mov edx, [ebp + nb332_VFtab] mov ecx, [esp + nb332_n1] lea ecx, [ecx + ecx*2] shl ecx, 2 ;# load all values we need movd mm4, [edx + ecx*4] movd mm5, [edx + ecx*4 + 4] movd mm6, [edx + ecx*4 + 8] movd mm7, [edx + ecx*4 + 12] mov ecx, [esp + nb332_n1+4] lea ecx, [ecx + ecx*2] shl ecx, 2 punpckldq mm4, [edx + ecx*4] punpckldq mm5, [edx + ecx*4 + 4] punpckldq mm6, [edx + ecx*4 + 8] punpckldq mm7, [edx + ecx*4 + 12] pfmul mm6, mm0 ;# mm6 = Geps pfmul mm7, mm2 ;# mm7 = Heps2 pfadd mm5, mm6 pfadd mm5, mm7 ;# mm5 = Fp pfmul mm7, [esp + nb332_two] ;# two*Heps2 pfadd mm7, mm6 pfadd mm7, mm5 ;# mm7=FF pfmul mm5, mm0 ;# mm5=eps*Fp pfadd mm5, mm4 ;# mm5= VV pfmul mm5, [esp + nb332_qqOH] ;# vcoul=qq*VV pfmul mm7, [esp + nb332_qqOH] ;# fijC=qq*FF ;# update vctot pfadd mm5, [esp + nb332_vctot] movq [esp + nb332_vctot], mm5 ;# change sign of fijC and multiply by rinv pxor mm4,mm4 pfsub mm4, mm7 pfmul mm4, [esp + nb332_tsc] pfmul mm4, mm1 ;# mm4 is total fscal (for the hydrogens) now ;# spread oxygen fscalar to both positions punpckldq mm3,mm3 ;# calc vectorial force for O movq mm0, [esp + nb332_dxO] movd mm1, [esp + nb332_dzO] pfmul mm0, mm3 pfmul mm1, mm3 ;# calc vectorial force for H's movq mm5, [esp + nb332_dxH] movq mm6, [esp + nb332_dyH] movq mm7, [esp + nb332_dzH] pfmul mm5, mm4 pfmul mm6, mm4 pfmul mm7, mm4 ;# update iO particle force movq mm2, [esp + nb332_fixO] movd mm3, [esp + nb332_fizO] pfadd mm2, mm0 pfadd mm3, mm1 movq [esp + nb332_fixO], mm2 movd [esp + nb332_fizO], mm3 ;# update iH forces movq mm2, [esp + nb332_fixH] movq mm3, [esp + nb332_fiyH] movq mm4, [esp + nb332_fizH] pfadd mm2, mm5 pfadd mm3, mm6 pfadd mm4, mm7 movq [esp + nb332_fixH], mm2 movq [esp + nb332_fiyH], mm3 movq [esp + nb332_fizH], mm4 ;# pack j forces from H in the same form as the oxygen force. pfacc mm5, mm6 ;# mm5(l)=fjx(H1+ h2) mm5(h)=fjy(H1+ h2) pfacc mm7, mm7 ;# mm7(l)=fjz(H1+ h2) pfadd mm0, mm5 ;# add up total force on j particle. pfadd mm1, mm7 ;# update j particle force movq mm2, [edi + eax*4] movd mm3, [edi + eax*4 + 8] pfsub mm2, mm0 pfsub mm3, mm1 movq [edi + eax*4], mm2 movd [edi + eax*4 +8], mm3 ;# interactions with j H1 movq mm0, [esi + eax*4 + 12] movd mm1, [esi + eax*4 + 20] ;# copy & expand to mm2-mm4 for the H interactions movq mm2, mm0 movq mm3, mm0 movq mm4, mm1 punpckldq mm2,mm2 punpckhdq mm3,mm3 punpckldq mm4,mm4 pfsubr mm0, [esp + nb332_ixO] pfsubr mm1, [esp + nb332_izO] movq [esp + nb332_dxO], mm0 pfmul mm0,mm0 movd [esp + nb332_dzO], mm1 pfmul mm1,mm1 pfacc mm0, mm1 pfadd mm0, mm1 ;# mm0=rsqO punpckldq mm2, mm2 punpckldq mm3, mm3 punpckldq mm4, mm4 ;# mm2-mm4 is jx-jz pfsubr mm2, [esp + nb332_ixH] pfsubr mm3, [esp + nb332_iyH] pfsubr mm4, [esp + nb332_izH] ;# mm2-mm4 is dxH-dzH movq [esp + nb332_dxH], mm2 movq [esp + nb332_dyH], mm3 movq [esp + nb332_dzH], mm4 pfmul mm2,mm2 pfmul mm3,mm3 pfmul mm4,mm4 pfadd mm3,mm2 pfadd mm3,mm4 ;# mm3=rsqH movq [esp + nb332_tmprsqH], mm3 pfrsqrt mm1,mm0 movq mm2,mm1 pfmul mm1,mm1 pfrsqit1 mm1,mm0 pfrcpit2 mm1,mm2 ;# mm1=invsqrt pfmul mm0, mm1 ;# mm0=rsq pfmul mm0, [esp + nb332_tsc] pf2iw mm4, mm0 movd [esp + nb332_n1], mm4 pi2fd mm4,mm4 pfsub mm0, mm4 ;# now mm0 is eps and mm4 n0 movq mm2, mm0 pfmul mm2, mm2 ;# mm0 is eps, mm2 eps2 ;# coulomb table mov edx, [ebp + nb332_VFtab] mov ecx, [esp + nb332_n1] lea ecx, [ecx + ecx*2] shl ecx, 2 ;# load all values we need movd mm4, [edx + ecx*4] movd mm5, [edx + ecx*4 + 4] movd mm6, [edx + ecx*4 + 8] movd mm7, [edx + ecx*4 + 12] pfmul mm6, mm0 ;# mm6 = Geps pfmul mm7, mm2 ;# mm7 = Heps2 pfadd mm5, mm6 pfadd mm5, mm7 ;# mm5 = Fp pfmul mm7, [esp + nb332_two] ;# two*Heps2 pfadd mm7, mm6 pfadd mm7, mm5 ;# mm7=FF pfmul mm5, mm0 ;# mm5=eps*Fp pfadd mm5, mm4 ;# mm5= VV pfmul mm5, [esp + nb332_qqOH] ;# vcoul=qq*VV pfmul mm7, [esp + nb332_qqOH] ;# fijC=qq*FF ;# update vctot directly, force is moved to mm3. pfadd mm5, [esp + nb332_vctot] movq [esp + nb332_vctot], mm5 pxor mm3, mm3 pfsub mm3, mm7 pfmul mm3, [esp + nb332_tsc] pfmul mm3, mm1 ;# mm3 is total fscal (for the oxygen) now movq mm0, [esp + nb332_tmprsqH] pfrsqrt mm1, mm0 pswapd mm0,mm0 pfrsqrt mm2, mm0 pswapd mm0,mm0 punpckldq mm1,mm2 ;# seeds are in mm1 now, and rsq in mm0. movq mm2, mm1 pfmul mm1,mm1 pfrsqit1 mm1,mm0 pfrcpit2 mm1,mm2 ;# mm1=invsqrt pfmul mm0,mm1 ;# mm0=r pfmul mm0, [esp + nb332_tsc] pf2iw mm4, mm0 movq [esp + nb332_n1], mm4 pi2fd mm4,mm4 pfsub mm0, mm4 ;# now mm0 is eps and mm4 n0 movq mm2, mm0 pfmul mm2, mm2 ;# mm0 is eps, mm2 eps2 ;# coulomb table mov edx, [ebp + nb332_VFtab] mov ecx, [esp + nb332_n1] lea ecx, [ecx + ecx*2] shl ecx, 2 ;# load all values we need movd mm4, [edx + ecx*4] movd mm5, [edx + ecx*4 + 4] movd mm6, [edx + ecx*4 + 8] movd mm7, [edx + ecx*4 + 12] mov ecx, [esp + nb332_n1+4] lea ecx, [ecx + ecx*2] shl ecx, 2 punpckldq mm4, [edx + ecx*4] punpckldq mm5, [edx + ecx*4 + 4] punpckldq mm6, [edx + ecx*4 + 8] punpckldq mm7, [edx + ecx*4 + 12] pfmul mm6, mm0 ;# mm6 = Geps pfmul mm7, mm2 ;# mm7 = Heps2 pfadd mm5, mm6 pfadd mm5, mm7 ;# mm5 = Fp pfmul mm7, [esp + nb332_two] ;# two*Heps2 pfadd mm7, mm6 pfadd mm7, mm5 ;# mm7=FF pfmul mm5, mm0 ;# mm5=eps*Fp pfadd mm5, mm4 ;# mm5= VV pfmul mm5, [esp + nb332_qqHH] ;# vcoul=qq*VV pfmul mm7, [esp + nb332_qqHH] ;# fijC=qq*FF ;# update vctot pfadd mm5, [esp + nb332_vctot] movq [esp + nb332_vctot], mm5 ;# change sign of fijC and multiply by rinv pxor mm4,mm4 pfsub mm4, mm7 pfmul mm4, [esp + nb332_tsc] pfmul mm4, mm1 ;# mm4 is total fscal (for the hydrogens) now ;# spread oxygen fscalar to both positions punpckldq mm3,mm3 ;# calc vectorial force for O movq mm0, [esp + nb332_dxO] movd mm1, [esp + nb332_dzO] pfmul mm0, mm3 pfmul mm1, mm3 ;# calc vectorial force for H's movq mm5, [esp + nb332_dxH] movq mm6, [esp + nb332_dyH] movq mm7, [esp + nb332_dzH] pfmul mm5, mm4 pfmul mm6, mm4 pfmul mm7, mm4 ;# update iO particle force movq mm2, [esp + nb332_fixO] movd mm3, [esp + nb332_fizO] pfadd mm2, mm0 pfadd mm3, mm1 movq [esp + nb332_fixO], mm2 movd [esp + nb332_fizO], mm3 ;# update iH forces movq mm2, [esp + nb332_fixH] movq mm3, [esp + nb332_fiyH] movq mm4, [esp + nb332_fizH] pfadd mm2, mm5 pfadd mm3, mm6 pfadd mm4, mm7 movq [esp + nb332_fixH], mm2 movq [esp + nb332_fiyH], mm3 movq [esp + nb332_fizH], mm4 ;# pack j forces from H in the same form as the oxygen force. pfacc mm5, mm6 ;# mm5(l)=fjx(H1+ h2) mm5(h)=fjy(H1+ h2) pfacc mm7, mm7 ;# mm7(l)=fjz(H1+ h2) pfadd mm0, mm5 ;# add up total force on j particle. pfadd mm1, mm7 ;# update j particle force movq mm2, [edi + eax*4 + 12] movd mm3, [edi + eax*4 + 20] pfsub mm2, mm0 pfsub mm3, mm1 movq [edi + eax*4 + 12], mm2 movd [edi + eax*4 + 20], mm3 ;# interactions with j H2 movq mm0, [esi + eax*4 + 24] movd mm1, [esi + eax*4 + 32] ;# copy & expand to mm2-mm4 for the H interactions movq mm2, mm0 movq mm3, mm0 movq mm4, mm1 punpckldq mm2,mm2 punpckhdq mm3,mm3 punpckldq mm4,mm4 pfsubr mm0, [esp + nb332_ixO] pfsubr mm1, [esp + nb332_izO] movq [esp + nb332_dxO], mm0 pfmul mm0,mm0 movd [esp + nb332_dzO], mm1 pfmul mm1,mm1 pfacc mm0, mm1 pfadd mm0, mm1 ;# mm0=rsqO punpckldq mm2, mm2 punpckldq mm3, mm3 punpckldq mm4, mm4 ;# mm2-mm4 is jx-jz pfsubr mm2, [esp + nb332_ixH] pfsubr mm3, [esp + nb332_iyH] pfsubr mm4, [esp + nb332_izH] ;# mm2-mm4 is dxH-dzH movq [esp + nb332_dxH], mm2 movq [esp + nb332_dyH], mm3 movq [esp + nb332_dzH], mm4 pfmul mm2,mm2 pfmul mm3,mm3 pfmul mm4,mm4 pfadd mm3,mm2 pfadd mm3,mm4 ;# mm3=rsqH movq [esp + nb332_tmprsqH], mm3 pfrsqrt mm1,mm0 movq mm2,mm1 pfmul mm1,mm1 pfrsqit1 mm1,mm0 pfrcpit2 mm1,mm2 ;# mm1=invsqrt pfmul mm0, mm1 pfmul mm0, [esp + nb332_tsc] pf2iw mm4, mm0 movd [esp + nb332_n1], mm4 pi2fd mm4,mm4 pfsub mm0, mm4 ;# now mm0 is eps and mm4 n0 movq mm2, mm0 pfmul mm2, mm2 ;# mm0 is eps, mm2 eps2 ;# coulomb table mov edx, [ebp + nb332_VFtab] mov ecx, [esp + nb332_n1] lea ecx, [ecx + ecx*2] shl ecx, 2 ;# load all values we need movd mm4, [edx + ecx*4] movd mm5, [edx + ecx*4 + 4] movd mm6, [edx + ecx*4 + 8] movd mm7, [edx + ecx*4 + 12] pfmul mm6, mm0 ;# mm6 = Geps pfmul mm7, mm2 ;# mm7 = Heps2 pfadd mm5, mm6 pfadd mm5, mm7 ;# mm5 = Fp pfmul mm7, [esp + nb332_two] ;# two*Heps2 pfadd mm7, mm6 pfadd mm7, mm5 ;# mm7=FF pfmul mm5, mm0 ;# mm5=eps*Fp pfadd mm5, mm4 ;# mm5= VV pfmul mm5, [esp + nb332_qqOH] ;# vcoul=qq*VV pfmul mm7, [esp + nb332_qqOH] ;# fijC=qq*FF ;# update vctot directly, use mm3 for fscal sum pfadd mm5, [esp + nb332_vctot] movq [esp + nb332_vctot], mm5 pxor mm3,mm3 pfsub mm3, mm7 pfmul mm3, [esp + nb332_tsc] pfmul mm3, mm1 ;# mm3 is total fscal (for the oxygen) now movq mm0, [esp + nb332_tmprsqH] pfrsqrt mm1, mm0 pswapd mm0,mm0 pfrsqrt mm2, mm0 pswapd mm0,mm0 punpckldq mm1,mm2 ;# seeds are in mm1 now, and rsq in mm0. movq mm2, mm1 pfmul mm1,mm1 pfrsqit1 mm1,mm0 pfrcpit2 mm1,mm2 ;# mm1=invsqrt pfmul mm0,mm1 ;# mm0=r pfmul mm0, [esp + nb332_tsc] pf2iw mm4, mm0 movq [esp + nb332_n1], mm4 pi2fd mm4,mm4 pfsub mm0, mm4 ;# now mm0 is eps and mm4 n0 movq mm2, mm0 pfmul mm2, mm2 ;# mm0 is eps, mm2 eps2 ;# coulomb table mov edx, [ebp + nb332_VFtab] mov ecx, [esp + nb332_n1] lea ecx, [ecx + ecx*2] shl ecx, 2 ;# load all values we need movd mm4, [edx + ecx*4] movd mm5, [edx + ecx*4 + 4] movd mm6, [edx + ecx*4 + 8] movd mm7, [edx + ecx*4 + 12] mov ecx, [esp + nb332_n1+4];# mm5 = Fp lea ecx, [ecx + ecx*2] shl ecx, 2 punpckldq mm4, [edx + ecx*4] punpckldq mm5, [edx + ecx*4 + 4] punpckldq mm6, [edx + ecx*4 + 8] punpckldq mm7, [edx + ecx*4 + 12] pfmul mm6, mm0 ;# mm6 = Geps pfmul mm7, mm2 ;# mm7 = Heps2 pfadd mm5, mm6 pfadd mm5, mm7 ;# mm5 = Fp pfmul mm7, [esp + nb332_two] ;# two*Heps2 pfadd mm7, mm6
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -