nb_kernel332_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,165 行 · 第 1/5 页
S
2,165 行
movaps xmm0, xmm1 movaps xmm2, xmm1 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb332_dxOO] mulps xmm1, [esp + nb332_dyOO] mulps xmm2, [esp + nb332_dzOO] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb332_fixO] addps xmm1, [esp + nb332_fiyO] addps xmm2, [esp + nb332_fizO] movaps [esp + nb332_fjxO], xmm3 movaps [esp + nb332_fjyO], xmm4 movaps [esp + nb332_fjzO], xmm5 movaps [esp + nb332_fixO], xmm0 movaps [esp + nb332_fiyO], xmm1 movaps [esp + nb332_fizO], xmm2 ;# O-H1 interaction movaps xmm0, [esp + nb332_rinvOH1] movaps xmm1, xmm0 mulps xmm1, [esp + nb332_rsqOH1] ;# xmm1=r mulps xmm1, [esp + nb332_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 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 + nb332_two] ;# two*Heps2 movaps xmm3, [esp + nb332_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 addps xmm5, [esp + nb332_vctot] movaps [esp + nb332_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb332_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb332_dxOH1] mulps xmm1, [esp + nb332_dyOH1] mulps xmm2, [esp + nb332_dzOH1] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb332_fixO] addps xmm1, [esp + nb332_fiyO] addps xmm2, [esp + nb332_fizO] movaps [esp + nb332_fjxH1], xmm3 movaps [esp + nb332_fjyH1], xmm4 movaps [esp + nb332_fjzH1], xmm5 movaps [esp + nb332_fixO], xmm0 movaps [esp + nb332_fiyO], xmm1 movaps [esp + nb332_fizO], xmm2 ;# O-H2 interaction movaps xmm0, [esp + nb332_rinvOH2] movaps xmm1, xmm0 mulps xmm1, [esp + nb332_rsqOH2] ;# xmm1=r mulps xmm1, [esp + nb332_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 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 + nb332_two] ;# two*Heps2 movaps xmm3, [esp + nb332_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 addps xmm5, [esp + nb332_vctot] movaps [esp + nb332_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb332_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb332_dxOH2] mulps xmm1, [esp + nb332_dyOH2] mulps xmm2, [esp + nb332_dzOH2] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb332_fixO] addps xmm1, [esp + nb332_fiyO] addps xmm2, [esp + nb332_fizO] movaps [esp + nb332_fjxH2], xmm3 movaps [esp + nb332_fjyH2], xmm4 movaps [esp + nb332_fjzH2], xmm5 movaps [esp + nb332_fixO], xmm0 movaps [esp + nb332_fiyO], xmm1 movaps [esp + nb332_fizO], xmm2 ;# H1-O interaction movaps xmm0, [esp + nb332_rinvH1O] movaps xmm1, xmm0 mulps xmm1, [esp + nb332_rsqH1O] ;# xmm1=r mulps xmm1, [esp + nb332_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 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 + nb332_two] ;# two*Heps2 movaps xmm3, [esp + nb332_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 addps xmm5, [esp + nb332_vctot] movaps [esp + nb332_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb332_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 movaps xmm3, [esp + nb332_fjxO] movaps xmm4, [esp + nb332_fjyO] movaps xmm5, [esp + nb332_fjzO] mulps xmm0, [esp + nb332_dxH1O] mulps xmm1, [esp + nb332_dyH1O] mulps xmm2, [esp + nb332_dzH1O] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb332_fixH1] addps xmm1, [esp + nb332_fiyH1] addps xmm2, [esp + nb332_fizH1] movaps [esp + nb332_fjxO], xmm3 movaps [esp + nb332_fjyO], xmm4 movaps [esp + nb332_fjzO], xmm5 movaps [esp + nb332_fixH1], xmm0 movaps [esp + nb332_fiyH1], xmm1 movaps [esp + nb332_fizH1], xmm2 ;# H1-H1 interaction movaps xmm0, [esp + nb332_rinvH1H1] movaps xmm1, xmm0 mulps xmm1, [esp + nb332_rsqH1H1] ;# xmm1=r mulps xmm1, [esp + nb332_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 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 + nb332_two] ;# two*Heps2 movaps xmm3, [esp + nb332_qqHH] 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 addps xmm5, [esp + nb332_vctot] movaps [esp + nb332_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb332_tsc] mulps xmm3, xmm0 subps xmm1, xmm3 movaps xmm0, xmm1 movaps xmm2, xmm1 movaps xmm3, [esp + nb332_fjxH1] movaps xmm4, [esp + nb332_fjyH1] movaps xmm5, [esp + nb332_fjzH1] mulps xmm0, [esp + nb332_dxH1H1] mulps xmm1, [esp + nb332_dyH1H1] mulps xmm2, [esp + nb332_dzH1H1] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb332_fixH1] addps xmm1, [esp + nb332_fiyH1] addps xmm2, [esp + nb332_fizH1] movaps [esp + nb332_fjxH1], xmm3 movaps [esp + nb332_fjyH1], xmm4 movaps [esp + nb332_fjzH1], xmm5 movaps [esp + nb332_fixH1], xmm0 movaps [esp + nb332_fiyH1], xmm1 movaps [esp + nb332_fizH1], xmm2 ;# H1-H2 interaction movaps xmm0, [esp + nb332_rinvH1H2] movaps xmm1, xmm0 mulps xmm1, [esp + nb332_rsqH1H2] ;# xmm1=r mulps xmm1, [esp + nb332_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 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
⌨️ 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?