nb_kernel232_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,146 行 · 第 1/5 页
S
2,146 行
movlps xmm6, [esi + ebx*4 + 12] movlps xmm7, [esi + ebx*4 + 24] movhps xmm2, [esi + ecx*4] movhps xmm3, [esi + ecx*4 + 12] movhps xmm4, [esi + ecx*4 + 24] movhps xmm5, [esi + edx*4] movhps xmm6, [esi + edx*4 + 12] movhps xmm7, [esi + edx*4 + 24] ;# current state: ;# xmm2= jxOa jyOa jxOc jyOc ;# xmm3= jxH1a jyH1a jxH1c jyH1c ;# xmm4= jxH2a jyH2a jxH2c jyH2c ;# xmm5= jxOb jyOb jxOd jyOd ;# xmm6= jxH1b jyH1b jxH1d jyH1d ;# xmm7= jxH2b jyH2b jxH2d jyH2d 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 + nb232_jxO], xmm0 movhlps xmm2, xmm6 ;# xmm2= jyOa jyOb jyOc jyOd movaps [esp + nb232_jyO], xmm2 movlhps xmm1, xmm3 movaps [esp + nb232_jxH1], xmm1 movhlps xmm3, xmm7 movaps xmm6, xmm4 movaps [esp + nb232_jyH1], xmm3 movlhps xmm4, xmm5 movaps [esp + nb232_jxH2], xmm4 movhlps xmm5, xmm6 movaps [esp + nb232_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 + nb232_jzO], xmm0 movaps [esp + nb232_jzH1], xmm1 movaps [esp + nb232_jzH2], xmm2 movaps xmm0, [esp + nb232_ixO] movaps xmm1, [esp + nb232_iyO] movaps xmm2, [esp + nb232_izO] movaps xmm3, [esp + nb232_ixO] movaps xmm4, [esp + nb232_iyO] movaps xmm5, [esp + nb232_izO] subps xmm0, [esp + nb232_jxO] subps xmm1, [esp + nb232_jyO] subps xmm2, [esp + nb232_jzO] subps xmm3, [esp + nb232_jxH1] subps xmm4, [esp + nb232_jyH1] subps xmm5, [esp + nb232_jzH1] movaps [esp + nb232_dxOO], xmm0 movaps [esp + nb232_dyOO], xmm1 movaps [esp + nb232_dzOO], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb232_dxOH1], xmm3 movaps [esp + nb232_dyOH1], xmm4 movaps [esp + nb232_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 + nb232_rsqOO], xmm0 movaps [esp + nb232_rsqOH1], xmm3 movaps xmm0, [esp + nb232_ixO] movaps xmm1, [esp + nb232_iyO] movaps xmm2, [esp + nb232_izO] movaps xmm3, [esp + nb232_ixH1] movaps xmm4, [esp + nb232_iyH1] movaps xmm5, [esp + nb232_izH1] subps xmm0, [esp + nb232_jxH2] subps xmm1, [esp + nb232_jyH2] subps xmm2, [esp + nb232_jzH2] subps xmm3, [esp + nb232_jxO] subps xmm4, [esp + nb232_jyO] subps xmm5, [esp + nb232_jzO] movaps [esp + nb232_dxOH2], xmm0 movaps [esp + nb232_dyOH2], xmm1 movaps [esp + nb232_dzOH2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb232_dxH1O], xmm3 movaps [esp + nb232_dyH1O], xmm4 movaps [esp + nb232_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 + nb232_rsqOH2], xmm0 movaps [esp + nb232_rsqH1O], xmm3 movaps xmm0, [esp + nb232_ixH1] movaps xmm1, [esp + nb232_iyH1] movaps xmm2, [esp + nb232_izH1] movaps xmm3, [esp + nb232_ixH1] movaps xmm4, [esp + nb232_iyH1] movaps xmm5, [esp + nb232_izH1] subps xmm0, [esp + nb232_jxH1] subps xmm1, [esp + nb232_jyH1] subps xmm2, [esp + nb232_jzH1] subps xmm3, [esp + nb232_jxH2] subps xmm4, [esp + nb232_jyH2] subps xmm5, [esp + nb232_jzH2] movaps [esp + nb232_dxH1H1], xmm0 movaps [esp + nb232_dyH1H1], xmm1 movaps [esp + nb232_dzH1H1], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb232_dxH1H2], xmm3 movaps [esp + nb232_dyH1H2], xmm4 movaps [esp + nb232_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 + nb232_rsqH1H1], xmm0 movaps [esp + nb232_rsqH1H2], xmm3 movaps xmm0, [esp + nb232_ixH2] movaps xmm1, [esp + nb232_iyH2] movaps xmm2, [esp + nb232_izH2] movaps xmm3, [esp + nb232_ixH2] movaps xmm4, [esp + nb232_iyH2] movaps xmm5, [esp + nb232_izH2] subps xmm0, [esp + nb232_jxO] subps xmm1, [esp + nb232_jyO] subps xmm2, [esp + nb232_jzO] subps xmm3, [esp + nb232_jxH1] subps xmm4, [esp + nb232_jyH1] subps xmm5, [esp + nb232_jzH1] movaps [esp + nb232_dxH2O], xmm0 movaps [esp + nb232_dyH2O], xmm1 movaps [esp + nb232_dzH2O], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb232_dxH2H1], xmm3 movaps [esp + nb232_dyH2H1], xmm4 movaps [esp + nb232_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 + nb232_rsqH2O], xmm0 movaps [esp + nb232_rsqH2H1], xmm4 movaps xmm0, [esp + nb232_ixH2] movaps xmm1, [esp + nb232_iyH2] movaps xmm2, [esp + nb232_izH2] subps xmm0, [esp + nb232_jxH2] subps xmm1, [esp + nb232_jyH2] subps xmm2, [esp + nb232_jzH2] movaps [esp + nb232_dxH2H2], xmm0 movaps [esp + nb232_dyH2H2], xmm1 movaps [esp + nb232_dzH2H2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 movaps [esp + nb232_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 + nb232_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 + nb232_half] ;# rinvH2H2 mulps xmm7, [esp + nb232_half] ;# rinvH2H1 movaps [esp + nb232_rinvH2H2], xmm3 movaps [esp + nb232_rinvH2H1], xmm7 rsqrtps xmm1, [esp + nb232_rsqOO] rsqrtps xmm5, [esp + nb232_rsqOH1] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb232_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb232_rsqOO] mulps xmm5, [esp + nb232_rsqOH1] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb232_half] mulps xmm7, [esp + nb232_half] movaps [esp + nb232_rinvOO], xmm3 movaps [esp + nb232_rinvOH1], xmm7 rsqrtps xmm1, [esp + nb232_rsqOH2] rsqrtps xmm5, [esp + nb232_rsqH1O] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb232_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb232_rsqOH2] mulps xmm5, [esp + nb232_rsqH1O] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb232_half] mulps xmm7, [esp + nb232_half] movaps [esp + nb232_rinvOH2], xmm3 movaps [esp + nb232_rinvH1O], xmm7 rsqrtps xmm1, [esp + nb232_rsqH1H1] rsqrtps xmm5, [esp + nb232_rsqH1H2] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb232_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb232_rsqH1H1] mulps xmm5, [esp + nb232_rsqH1H2] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb232_half] mulps xmm7, [esp + nb232_half] movaps [esp + nb232_rinvH1H1], xmm3 movaps [esp + nb232_rinvH1H2], xmm7 rsqrtps xmm1, [esp + nb232_rsqH2O] movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [esp + nb232_three] mulps xmm1, [esp + nb232_rsqH2O] subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [esp + nb232_half] movaps [esp + nb232_rinvH2O], xmm3 ;# start with OO interaction - first the table LJ part movaps xmm0, [esp + nb232_rinvOO] movaps xmm1, xmm0 mulps xmm1, [esp + nb232_rsqOO] ;# xmm1=r mulps xmm1, [esp + nb232_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, 3 pslld mm7, 3 movd mm0, eax movd mm1, ebx movd mm2, ecx movd mm3, edx movd eax, mm6 psrlq mm6, 32 movd ecx, mm7 psrlq mm7, 32 movd ebx, mm6 movd edx, mm7 mov esi, [ebp + nb232_VFtab] ;# dispersion movlps xmm5, [esi + eax*4] movlps xmm7, [esi + ecx*4] movhps xmm5, [esi + ebx*4] movhps xmm7, [esi + edx*4] ;# got half 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 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 + nb232_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 + nb232_c6] mulps xmm7, xmm4 ;# fijD mulps xmm5, xmm4 ;# Vvdw6 movaps [esp + nb232_fstmp], xmm7 addps xmm5, [esp + nb232_Vvdwtot] movaps [esp + nb232_Vvdwtot], xmm5 ;# repulsion 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 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 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 + nb232_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 + nb232_c12] mulps xmm7, xmm4 ;# fijR mulps xmm5, xmm4 ;# Vvdw12 addps xmm7, [esp + nb232_fstmp] mulps xmm7, [esp + nb232_tsc] addps xmm5, [esp + nb232_Vvdwtot] movaps [esp + nb232_Vvdwtot], xmm5 movaps xmm0, [esp + nb232_rinvOO] movaps xmm2, xmm0 ;# xmm7=rinv movaps xmm5, [esp + nb232_krf] mulps xmm5, [esp + nb232_rsqOO] ;# xmm5=krsq movaps xmm6, xmm5 addps xmm6, xmm2 ;# xmm6=rinv+ krsq subps xmm6, [esp + nb232_crf] mulps xmm6, [esp + nb232_qqOO] ;# xmm6=voul=qq*(rinv+ krsq-crf) mulps xmm5, [esp + nb232_two] subps xmm2, xmm5 ;# xmm2=rinv-2*krsq mulps xmm2, [esp + nb232_qqOO] ;# xmm2 = coul part of fscal addps xmm6, [esp + nb232_vctot] ;# local vctot summation variable mulps xmm2, xmm0 subps xmm2, xmm7 mulps xmm2, xmm0 movaps xmm0, xmm2 movaps xmm1, xmm2 movd eax, mm0 movd ebx, mm1 movd ecx, mm2 movd edx, mm3 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb232_dxOO]
⌨️ 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?