nb_kernel304_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,258 行 · 第 1/5 页
S
2,258 行
shufps xmm0, xmm3, 204 ;# constant 11001100 shufps xmm1, xmm4, 204 ;# constant 11001100 shufps xmm2, xmm5, 204 ;# constant 11001100 movaps [esp + nb304_jzH1], xmm0 movaps [esp + nb304_jzH2], xmm1 movaps [esp + nb304_jzM], xmm2 movaps xmm0, [esp + nb304_ixH1] movaps xmm1, [esp + nb304_iyH1] movaps xmm2, [esp + nb304_izH1] movaps xmm3, [esp + nb304_ixH1] movaps xmm4, [esp + nb304_iyH1] movaps xmm5, [esp + nb304_izH1] subps xmm0, [esp + nb304_jxH1] subps xmm1, [esp + nb304_jyH1] subps xmm2, [esp + nb304_jzH1] subps xmm3, [esp + nb304_jxH2] subps xmm4, [esp + nb304_jyH2] subps xmm5, [esp + nb304_jzH2] movaps [esp + nb304_dxH1H1], xmm0 movaps [esp + nb304_dyH1H1], xmm1 movaps [esp + nb304_dzH1H1], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb304_dxH1H2], xmm3 movaps [esp + nb304_dyH1H2], xmm4 movaps [esp + nb304_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 + nb304_rsqH1H1], xmm0 movaps [esp + nb304_rsqH1H2], xmm3 movaps xmm0, [esp + nb304_ixH1] movaps xmm1, [esp + nb304_iyH1] movaps xmm2, [esp + nb304_izH1] movaps xmm3, [esp + nb304_ixH2] movaps xmm4, [esp + nb304_iyH2] movaps xmm5, [esp + nb304_izH2] subps xmm0, [esp + nb304_jxM] subps xmm1, [esp + nb304_jyM] subps xmm2, [esp + nb304_jzM] subps xmm3, [esp + nb304_jxH1] subps xmm4, [esp + nb304_jyH1] subps xmm5, [esp + nb304_jzH1] movaps [esp + nb304_dxH1M], xmm0 movaps [esp + nb304_dyH1M], xmm1 movaps [esp + nb304_dzH1M], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb304_dxH2H1], xmm3 movaps [esp + nb304_dyH2H1], xmm4 movaps [esp + nb304_dzH2H1], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm3, xmm4 addps xmm3, xmm5 movaps [esp + nb304_rsqH1M], xmm0 movaps [esp + nb304_rsqH2H1], xmm3 movaps xmm0, [esp + nb304_ixH2] movaps xmm1, [esp + nb304_iyH2] movaps xmm2, [esp + nb304_izH2] movaps xmm3, [esp + nb304_ixH2] movaps xmm4, [esp + nb304_iyH2] movaps xmm5, [esp + nb304_izH2] subps xmm0, [esp + nb304_jxH2] subps xmm1, [esp + nb304_jyH2] subps xmm2, [esp + nb304_jzH2] subps xmm3, [esp + nb304_jxM] subps xmm4, [esp + nb304_jyM] subps xmm5, [esp + nb304_jzM] movaps [esp + nb304_dxH2H2], xmm0 movaps [esp + nb304_dyH2H2], xmm1 movaps [esp + nb304_dzH2H2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb304_dxH2M], xmm3 movaps [esp + nb304_dyH2M], xmm4 movaps [esp + nb304_dzH2M], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm3, xmm4 addps xmm3, xmm5 movaps [esp + nb304_rsqH2H2], xmm0 movaps [esp + nb304_rsqH2M], xmm3 movaps xmm0, [esp + nb304_ixM] movaps xmm1, [esp + nb304_iyM] movaps xmm2, [esp + nb304_izM] movaps xmm3, [esp + nb304_ixM] movaps xmm4, [esp + nb304_iyM] movaps xmm5, [esp + nb304_izM] subps xmm0, [esp + nb304_jxH1] subps xmm1, [esp + nb304_jyH1] subps xmm2, [esp + nb304_jzH1] subps xmm3, [esp + nb304_jxH2] subps xmm4, [esp + nb304_jyH2] subps xmm5, [esp + nb304_jzH2] movaps [esp + nb304_dxMH1], xmm0 movaps [esp + nb304_dyMH1], xmm1 movaps [esp + nb304_dzMH1], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb304_dxMH2], xmm3 movaps [esp + nb304_dyMH2], xmm4 movaps [esp + nb304_dzMH2], xmm5 mulps xmm3, xmm3 mulps xmm4, xmm4 mulps xmm5, xmm5 addps xmm0, xmm1 addps xmm0, xmm2 addps xmm4, xmm3 addps xmm4, xmm5 movaps [esp + nb304_rsqMH1], xmm0 movaps [esp + nb304_rsqMH2], xmm4 movaps xmm0, [esp + nb304_ixM] movaps xmm1, [esp + nb304_iyM] movaps xmm2, [esp + nb304_izM] subps xmm0, [esp + nb304_jxM] subps xmm1, [esp + nb304_jyM] subps xmm2, [esp + nb304_jzM] movaps [esp + nb304_dxMM], xmm0 movaps [esp + nb304_dyMM], xmm1 movaps [esp + nb304_dzMM], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 movaps [esp + nb304_rsqMM], 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 + nb304_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 + nb304_half] ;# rinvMM mulps xmm7, [esp + nb304_half] ;# rinvMH2 movaps [esp + nb304_rinvMM], xmm3 movaps [esp + nb304_rinvMH2], xmm7 rsqrtps xmm1, [esp + nb304_rsqH1H1] rsqrtps xmm5, [esp + nb304_rsqH1H2] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb304_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb304_rsqH1H1] mulps xmm5, [esp + nb304_rsqH1H2] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb304_half] mulps xmm7, [esp + nb304_half] movaps [esp + nb304_rinvH1H1], xmm3 movaps [esp + nb304_rinvH1H2], xmm7 rsqrtps xmm1, [esp + nb304_rsqH1M] rsqrtps xmm5, [esp + nb304_rsqH2H1] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb304_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb304_rsqH1M] mulps xmm5, [esp + nb304_rsqH2H1] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb304_half] mulps xmm7, [esp + nb304_half] movaps [esp + nb304_rinvH1M], xmm3 movaps [esp + nb304_rinvH2H1], xmm7 rsqrtps xmm1, [esp + nb304_rsqH2H2] rsqrtps xmm5, [esp + nb304_rsqH2M] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb304_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb304_rsqH2H2] mulps xmm5, [esp + nb304_rsqH2M] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb304_half] mulps xmm7, [esp + nb304_half] movaps [esp + nb304_rinvH2H2], xmm3 movaps [esp + nb304_rinvH2M], xmm7 rsqrtps xmm1, [esp + nb304_rsqMH1] movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [esp + nb304_three] mulps xmm1, [esp + nb304_rsqMH1] subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [esp + nb304_half] movaps [esp + nb304_rinvMH1], xmm3 ;# start with H1-H1 interaction movaps xmm0, [esp + nb304_rinvH1H1] movaps xmm1, xmm0 mulps xmm1, [esp + nb304_rsqH1H1] ;# xmm1=r mulps xmm1, [esp + nb304_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 + nb304_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 + nb304_two] ;# two*Heps2 movaps xmm3, [esp + nb304_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 ;# increment vcoul - then we can get rid of mm5 ;# update vctot addps xmm5, [esp + nb304_vctot] xorps xmm2, xmm2 movaps [esp + nb304_vctot], xmm5 mulps xmm3, [esp + nb304_tsc] subps xmm2, xmm3 mulps xmm0, xmm2 movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb304_dxH1H1] mulps xmm1, [esp + nb304_dyH1H1] mulps xmm2, [esp + nb304_dzH1H1] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb304_fixH1] addps xmm1, [esp + nb304_fiyH1] addps xmm2, [esp + nb304_fizH1] movaps [esp + nb304_fjxH1], xmm3 movaps [esp + nb304_fjyH1], xmm4 movaps [esp + nb304_fjzH1], xmm5 movaps [esp + nb304_fixH1], xmm0 movaps [esp + nb304_fiyH1], xmm1 movaps [esp + nb304_fizH1], xmm2 ;# H1-H2 interaction movaps xmm0, [esp + nb304_rinvH1H2] movaps xmm1, xmm0 mulps xmm1, [esp + nb304_rsqH1H2] ;# xmm1=r mulps xmm1, [esp + nb304_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 + nb304_two] ;# two*Heps2 movaps xmm3, [esp + nb304_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 + nb304_vctot] movaps [esp + nb304_vctot], xmm5 xorps xmm1, xmm1 mulps xmm3, [esp + nb304_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 + nb304_dxH1H2] mulps xmm1, [esp + nb304_dyH1H2] mulps xmm2, [esp + nb304_dzH1H2] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb304_fixH1] addps xmm1, [esp + nb304_fiyH1] addps xmm2, [esp + nb304_fizH1] movaps [esp + nb304_fjxH2], xmm3 movaps [esp + nb304_fjyH2], xmm4 movaps [esp + nb304_fjzH2], xmm5 movaps [esp + nb304_fixH1], xmm0 movaps [esp + nb304_fiyH1], xmm1 movaps [esp + nb304_fizH1], xmm2 ;# H1-M interaction movaps xmm0, [esp + nb304_rinvH1M] movaps xmm1, xmm0 mulps xmm1, [esp + nb304_rsqH1M] ;# xmm1=r mulps xmm1, [esp + nb304_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
⌨️ 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?