nb_kernel312_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,216 行 · 第 1/5 页
S
2,216 行
movaps xmm0, xmm2 movaps xmm1, xmm3 unpcklps xmm0, xmm5 ;# xmm0= jxOa jxOb jyOa jyOb unpcklps xmm1, xmm6 ;# xmm1= jxH1a jxH1b jyH1a jyH1b unpckhps xmm2, xmm5 ;# xmm2= jxOc jxOd jyOc jyOd unpckhps xmm3, xmm6 ;# xmm3= jxH1c jxH1d jyH1c jyH1d movaps xmm5, xmm4 movaps xmm6, xmm0 unpcklps xmm4, xmm7 ;# xmm4= jxH2a jxH2b jyH2a jyH2b unpckhps xmm5, xmm7 ;# xmm5= jxH2c jxH2d jyH2c jyH2d movaps xmm7, xmm1 movlhps xmm0, xmm2 ;# xmm0= jxOa jxOb jxOc jxOd movaps [esp + nb312_jxO], xmm0 movhlps xmm2, xmm6 ;# xmm2= jyOa jyOb jyOc jyOd movaps [esp + nb312_jyO], xmm2 movlhps xmm1, xmm3 movaps [esp + nb312_jxH1], xmm1 movhlps xmm3, xmm7 movaps xmm6, xmm4 movaps [esp + nb312_jyH1], xmm3 movlhps xmm4, xmm5 movaps [esp + nb312_jxH2], xmm4 movhlps xmm5, xmm6 movaps [esp + nb312_jyH2], xmm5 movss xmm0, [esi + eax*4 + 8] movss xmm1, [esi + eax*4 + 20] movss xmm2, [esi + eax*4 + 32] movss xmm3, [esi + ecx*4 + 8] movss xmm4, [esi + ecx*4 + 20] movss xmm5, [esi + ecx*4 + 32] movhps xmm0, [esi + ebx*4 + 4] movhps xmm1, [esi + ebx*4 + 16] movhps xmm2, [esi + ebx*4 + 28] movhps xmm3, [esi + edx*4 + 4] movhps xmm4, [esi + edx*4 + 16] movhps xmm5, [esi + edx*4 + 28] shufps xmm0, xmm3, 204 ;# constant 11001100 shufps xmm1, xmm4, 204 ;# constant 11001100 shufps xmm2, xmm5, 204 ;# constant 11001100 movaps [esp + nb312_jzO], xmm0 movaps [esp + nb312_jzH1], xmm1 movaps [esp + nb312_jzH2], xmm2 movaps xmm0, [esp + nb312_ixO] movaps xmm1, [esp + nb312_iyO] movaps xmm2, [esp + nb312_izO] movaps xmm3, [esp + nb312_ixO] movaps xmm4, [esp + nb312_iyO] movaps xmm5, [esp + nb312_izO] subps xmm0, [esp + nb312_jxO] subps xmm1, [esp + nb312_jyO] subps xmm2, [esp + nb312_jzO] subps xmm3, [esp + nb312_jxH1] subps xmm4, [esp + nb312_jyH1] subps xmm5, [esp + nb312_jzH1] movaps [esp + nb312_dxOO], xmm0 movaps [esp + nb312_dyOO], xmm1 movaps [esp + nb312_dzOO], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb312_dxOH1], xmm3 movaps [esp + nb312_dyOH1], xmm4 movaps [esp + nb312_dzOH1], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm3, xmm4 addps xmm3, xmm5 movaps [esp + nb312_rsqOO], xmm0 movaps [esp + nb312_rsqOH1], xmm3 movaps xmm0, [esp + nb312_ixO] movaps xmm1, [esp + nb312_iyO] movaps xmm2, [esp + nb312_izO] movaps xmm3, [esp + nb312_ixH1] movaps xmm4, [esp + nb312_iyH1] movaps xmm5, [esp + nb312_izH1] subps xmm0, [esp + nb312_jxH2] subps xmm1, [esp + nb312_jyH2] subps xmm2, [esp + nb312_jzH2] subps xmm3, [esp + nb312_jxO] subps xmm4, [esp + nb312_jyO] subps xmm5, [esp + nb312_jzO] movaps [esp + nb312_dxOH2], xmm0 movaps [esp + nb312_dyOH2], xmm1 movaps [esp + nb312_dzOH2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb312_dxH1O], xmm3 movaps [esp + nb312_dyH1O], xmm4 movaps [esp + nb312_dzH1O], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm3, xmm4 addps xmm3, xmm5 movaps [esp + nb312_rsqOH2], xmm0 movaps [esp + nb312_rsqH1O], xmm3 movaps xmm0, [esp + nb312_ixH1] movaps xmm1, [esp + nb312_iyH1] movaps xmm2, [esp + nb312_izH1] movaps xmm3, [esp + nb312_ixH1] movaps xmm4, [esp + nb312_iyH1] movaps xmm5, [esp + nb312_izH1] subps xmm0, [esp + nb312_jxH1] subps xmm1, [esp + nb312_jyH1] subps xmm2, [esp + nb312_jzH1] subps xmm3, [esp + nb312_jxH2] subps xmm4, [esp + nb312_jyH2] subps xmm5, [esp + nb312_jzH2] movaps [esp + nb312_dxH1H1], xmm0 movaps [esp + nb312_dyH1H1], xmm1 movaps [esp + nb312_dzH1H1], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb312_dxH1H2], xmm3 movaps [esp + nb312_dyH1H2], xmm4 movaps [esp + nb312_dzH1H2], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm3, xmm4 addps xmm3, xmm5 movaps [esp + nb312_rsqH1H1], xmm0 movaps [esp + nb312_rsqH1H2], xmm3 movaps xmm0, [esp + nb312_ixH2] movaps xmm1, [esp + nb312_iyH2] movaps xmm2, [esp + nb312_izH2] movaps xmm3, [esp + nb312_ixH2] movaps xmm4, [esp + nb312_iyH2] movaps xmm5, [esp + nb312_izH2] subps xmm0, [esp + nb312_jxO] subps xmm1, [esp + nb312_jyO] subps xmm2, [esp + nb312_jzO] subps xmm3, [esp + nb312_jxH1] subps xmm4, [esp + nb312_jyH1] subps xmm5, [esp + nb312_jzH1] movaps [esp + nb312_dxH2O], xmm0 movaps [esp + nb312_dyH2O], xmm1 movaps [esp + nb312_dzH2O], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb312_dxH2H1], xmm3 movaps [esp + nb312_dyH2H1], xmm4 movaps [esp + nb312_dzH2H1], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm4, xmm3 addps xmm4, xmm5 movaps [esp + nb312_rsqH2O], xmm0 movaps [esp + nb312_rsqH2H1], xmm4 movaps xmm0, [esp + nb312_ixH2] movaps xmm1, [esp + nb312_iyH2] movaps xmm2, [esp + nb312_izH2] subps xmm0, [esp + nb312_jxH2] subps xmm1, [esp + nb312_jyH2] subps xmm2, [esp + nb312_jzH2] movaps [esp + nb312_dxH2H2], xmm0 movaps [esp + nb312_dyH2H2], xmm1 movaps [esp + nb312_dzH2H2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 movaps [esp + nb312_rsqH2H2], xmm0 ;# start doing invsqrt use rsq values in xmm0, xmm4 rsqrtps xmm1, xmm0 rsqrtps xmm5, xmm4 movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb312_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 + nb312_half] ;# rinvH2H2 mulps xmm7, [esp + nb312_half] ;# rinvH2H1 movaps [esp + nb312_rinvH2H2], xmm3 movaps [esp + nb312_rinvH2H1], xmm7 rsqrtps xmm1, [esp + nb312_rsqOO] rsqrtps xmm5, [esp + nb312_rsqOH1] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb312_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb312_rsqOO] mulps xmm5, [esp + nb312_rsqOH1] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb312_half] mulps xmm7, [esp + nb312_half] movaps [esp + nb312_rinvOO], xmm3 movaps [esp + nb312_rinvOH1], xmm7 rsqrtps xmm1, [esp + nb312_rsqOH2] rsqrtps xmm5, [esp + nb312_rsqH1O] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb312_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb312_rsqOH2] mulps xmm5, [esp + nb312_rsqH1O] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb312_half] mulps xmm7, [esp + nb312_half] movaps [esp + nb312_rinvOH2], xmm3 movaps [esp + nb312_rinvH1O], xmm7 rsqrtps xmm1, [esp + nb312_rsqH1H1] rsqrtps xmm5, [esp + nb312_rsqH1H2] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb312_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb312_rsqH1H1] mulps xmm5, [esp + nb312_rsqH1H2] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb312_half] mulps xmm7, [esp + nb312_half] movaps [esp + nb312_rinvH1H1], xmm3 movaps [esp + nb312_rinvH1H2], xmm7 rsqrtps xmm1, [esp + nb312_rsqH2O] movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [esp + nb312_three] mulps xmm1, [esp + nb312_rsqH2O] subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [esp + nb312_half] movaps [esp + nb312_rinvH2O], xmm3 ;# start with OO interaction movaps xmm0, [esp + nb312_rinvOO] movaps xmm1, xmm0 mulps xmm1, [esp + nb312_rsqOO] ;# xmm1=r mulps xmm1, [esp + nb312_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 mm0, eax movd mm1, ebx movd mm2, ecx movd mm3, edx mov esi, [ebp + nb312_VFtab] movd eax, mm6 psrlq mm6, 32 movd ecx, mm7 psrlq mm7, 32 movd ebx, mm6 movd edx, mm7 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 + nb312_two] ;# two*Heps2 movaps xmm3, [esp + nb312_qqOO] 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 mm5 contains vcoul and mm3 fijC ;# increment vcoul - then we can get rid of mm5 ;# update vctot addps xmm5, [esp + nb312_vctot] movaps [esp + nb312_vctot], xmm5 mulps xmm3, [esp + nb312_tsc] ;# start doing lj movaps xmm2, xmm0 mulps xmm2, xmm2 movaps xmm1, xmm2 mulps xmm1, xmm2 mulps xmm1, xmm2 ;# xmm1=rinvsix movaps xmm2, xmm1 mulps xmm2, xmm2 ;# xmm2=rinvtwelve mulps xmm1, [esp + nb312_c6] mulps xmm2, [esp + nb312_c12] movaps xmm4, xmm2 subps xmm4, xmm1 addps xmm4, [esp + nb312_Vvdwtot] mulps xmm1, [esp + nb312_six] mulps xmm2, [esp + nb312_twelve] movaps [esp + nb312_Vvdwtot], xmm4 subps xmm2, xmm1 mulps xmm2, xmm0 subps xmm2, xmm3 mulps xmm0, xmm2 movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb312_dxOO] mulps xmm1, [esp + nb312_dyOO] mulps xmm2, [esp + nb312_dzOO] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb312_fixO] addps xmm1, [esp + nb312_fiyO] addps xmm2, [esp + nb312_fizO] movaps [esp + nb312_fjxO], xmm3 movaps [esp + nb312_fjyO], xmm4 movaps [esp + nb312_fjzO], xmm5 movaps [esp + nb312_fixO], xmm0 movaps [esp + nb312_fiyO], xmm1 movaps [esp + nb312_fizO], xmm2 ;# O-H1 interaction movaps xmm0, [esp + nb312_rinvOH1] movaps xmm1, xmm0 mulps xmm1, [esp + nb312_rsqOH1] ;# xmm1=r mulps xmm1, [esp + nb312_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 eax, mm6 psrlq mm6, 32 movd ecx, mm7 psrlq mm7, 32 movd ebx, mm6 movd edx, mm7 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 + nb312_two] ;# two*Heps2 movaps xmm3, [esp + nb312_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 mm5 contains vcoul and mm3 fijC
⌨️ 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?