nb_kernel104_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,174 行 · 第 1/5 页
S
2,174 行
movss xmm0, [esi + eax*4 + 20] movss xmm1, [esi + eax*4 + 32] movss xmm2, [esi + eax*4 + 44] movss xmm3, [esi + ecx*4 + 20] movss xmm4, [esi + ecx*4 + 32] movss xmm5, [esi + ecx*4 + 44] movhps xmm0, [esi + ebx*4 + 16] movhps xmm1, [esi + ebx*4 + 28] movhps xmm2, [esi + ebx*4 + 40] movhps xmm3, [esi + edx*4 + 16] movhps xmm4, [esi + edx*4 + 28] movhps xmm5, [esi + edx*4 + 40] shufps xmm0, xmm3, 204 ;# constant 11001100 shufps xmm1, xmm4, 204 ;# constant 11001100 shufps xmm2, xmm5, 204 ;# constant 11001100 movaps [esp + nb104_jzH1], xmm0 movaps [esp + nb104_jzH2], xmm1 movaps [esp + nb104_jzM], xmm2 movaps xmm0, [esp + nb104_ixH1] movaps xmm1, [esp + nb104_iyH1] movaps xmm2, [esp + nb104_izH1] movaps xmm3, [esp + nb104_ixH1] movaps xmm4, [esp + nb104_iyH1] movaps xmm5, [esp + nb104_izH1] subps xmm0, [esp + nb104_jxH1] subps xmm1, [esp + nb104_jyH1] subps xmm2, [esp + nb104_jzH1] subps xmm3, [esp + nb104_jxH2] subps xmm4, [esp + nb104_jyH2] subps xmm5, [esp + nb104_jzH2] movaps [esp + nb104_dxH1H1], xmm0 movaps [esp + nb104_dyH1H1], xmm1 movaps [esp + nb104_dzH1H1], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb104_dxH1H2], xmm3 movaps [esp + nb104_dyH1H2], xmm4 movaps [esp + nb104_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 + nb104_rsqH1H1], xmm0 movaps [esp + nb104_rsqH1H2], xmm3 movaps xmm0, [esp + nb104_ixH1] movaps xmm1, [esp + nb104_iyH1] movaps xmm2, [esp + nb104_izH1] movaps xmm3, [esp + nb104_ixH2] movaps xmm4, [esp + nb104_iyH2] movaps xmm5, [esp + nb104_izH2] subps xmm0, [esp + nb104_jxM] subps xmm1, [esp + nb104_jyM] subps xmm2, [esp + nb104_jzM] subps xmm3, [esp + nb104_jxH1] subps xmm4, [esp + nb104_jyH1] subps xmm5, [esp + nb104_jzH1] movaps [esp + nb104_dxH1M], xmm0 movaps [esp + nb104_dyH1M], xmm1 movaps [esp + nb104_dzH1M], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb104_dxH2H1], xmm3 movaps [esp + nb104_dyH2H1], xmm4 movaps [esp + nb104_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 + nb104_rsqH1M], xmm0 movaps [esp + nb104_rsqH2H1], xmm3 movaps xmm0, [esp + nb104_ixH2] movaps xmm1, [esp + nb104_iyH2] movaps xmm2, [esp + nb104_izH2] movaps xmm3, [esp + nb104_ixH2] movaps xmm4, [esp + nb104_iyH2] movaps xmm5, [esp + nb104_izH2] subps xmm0, [esp + nb104_jxH2] subps xmm1, [esp + nb104_jyH2] subps xmm2, [esp + nb104_jzH2] subps xmm3, [esp + nb104_jxM] subps xmm4, [esp + nb104_jyM] subps xmm5, [esp + nb104_jzM] movaps [esp + nb104_dxH2H2], xmm0 movaps [esp + nb104_dyH2H2], xmm1 movaps [esp + nb104_dzH2H2], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb104_dxH2M], xmm3 movaps [esp + nb104_dyH2M], xmm4 movaps [esp + nb104_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 + nb104_rsqH2H2], xmm0 movaps [esp + nb104_rsqH2M], xmm3 movaps xmm0, [esp + nb104_ixM] movaps xmm1, [esp + nb104_iyM] movaps xmm2, [esp + nb104_izM] movaps xmm3, [esp + nb104_ixM] movaps xmm4, [esp + nb104_iyM] movaps xmm5, [esp + nb104_izM] subps xmm0, [esp + nb104_jxH1] subps xmm1, [esp + nb104_jyH1] subps xmm2, [esp + nb104_jzH1] subps xmm3, [esp + nb104_jxH2] subps xmm4, [esp + nb104_jyH2] subps xmm5, [esp + nb104_jzH2] movaps [esp + nb104_dxMH1], xmm0 movaps [esp + nb104_dyMH1], xmm1 movaps [esp + nb104_dzMH1], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 movaps [esp + nb104_dxMH2], xmm3 movaps [esp + nb104_dyMH2], xmm4 movaps [esp + nb104_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 + nb104_rsqMH1], xmm0 movaps [esp + nb104_rsqMH2], xmm4 movaps xmm0, [esp + nb104_ixM] movaps xmm1, [esp + nb104_iyM] movaps xmm2, [esp + nb104_izM] subps xmm0, [esp + nb104_jxM] subps xmm1, [esp + nb104_jyM] subps xmm2, [esp + nb104_jzM] movaps [esp + nb104_dxMM], xmm0 movaps [esp + nb104_dyMM], xmm1 movaps [esp + nb104_dzMM], xmm2 mulps xmm0, xmm0 mulps xmm1, xmm1 mulps xmm2, xmm2 addps xmm0, xmm1 addps xmm0, xmm2 movaps [esp + nb104_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 + nb104_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 + nb104_half] ;# rinvMM mulps xmm7, [esp + nb104_half] ;# rinvMH2 movaps [esp + nb104_rinvMM], xmm3 movaps [esp + nb104_rinvMH2], xmm7 rsqrtps xmm1, [esp + nb104_rsqH1H1] rsqrtps xmm5, [esp + nb104_rsqH1H2] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb104_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb104_rsqH1H1] mulps xmm5, [esp + nb104_rsqH1H2] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb104_half] mulps xmm7, [esp + nb104_half] movaps [esp + nb104_rinvH1H1], xmm3 movaps [esp + nb104_rinvH1H2], xmm7 rsqrtps xmm1, [esp + nb104_rsqH1M] rsqrtps xmm5, [esp + nb104_rsqH2H1] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb104_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb104_rsqH1M] mulps xmm5, [esp + nb104_rsqH2H1] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb104_half] mulps xmm7, [esp + nb104_half] movaps [esp + nb104_rinvH1M], xmm3 movaps [esp + nb104_rinvH2H1], xmm7 rsqrtps xmm1, [esp + nb104_rsqH2H2] rsqrtps xmm5, [esp + nb104_rsqH2M] movaps xmm2, xmm1 movaps xmm6, xmm5 mulps xmm1, xmm1 mulps xmm5, xmm5 movaps xmm3, [esp + nb104_three] movaps xmm7, xmm3 mulps xmm1, [esp + nb104_rsqH2H2] mulps xmm5, [esp + nb104_rsqH2M] subps xmm3, xmm1 subps xmm7, xmm5 mulps xmm3, xmm2 mulps xmm7, xmm6 mulps xmm3, [esp + nb104_half] mulps xmm7, [esp + nb104_half] movaps [esp + nb104_rinvH2H2], xmm3 movaps [esp + nb104_rinvH2M], xmm7 rsqrtps xmm1, [esp + nb104_rsqMH1] movaps xmm2, xmm1 mulps xmm1, xmm1 movaps xmm3, [esp + nb104_three] mulps xmm1, [esp + nb104_rsqMH1] subps xmm3, xmm1 mulps xmm3, xmm2 mulps xmm3, [esp + nb104_half] movaps [esp + nb104_rinvMH1], xmm3 ;# start with H1-H1 interaction movaps xmm0, [esp + nb104_rinvH1H1] movaps xmm7, xmm0 mulps xmm0, xmm0 mulps xmm7, [esp + nb104_qqHH] mulps xmm0, xmm7 addps xmm7, [esp + nb104_vctot] movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb104_dxH1H1] mulps xmm1, [esp + nb104_dyH1H1] mulps xmm2, [esp + nb104_dzH1H1] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb104_fixH1] addps xmm1, [esp + nb104_fiyH1] addps xmm2, [esp + nb104_fizH1] movaps [esp + nb104_fjxH1], xmm3 movaps [esp + nb104_fjyH1], xmm4 movaps [esp + nb104_fjzH1], xmm5 movaps [esp + nb104_fixH1], xmm0 movaps [esp + nb104_fiyH1], xmm1 movaps [esp + nb104_fizH1], xmm2 ;# H1-H2 interaction movaps xmm0, [esp + nb104_rinvH1H2] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb104_qqHH] mulps xmm0, xmm1 ;# fs H1-H2 addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb104_dxH1H2] mulps xmm1, [esp + nb104_dyH1H2] mulps xmm2, [esp + nb104_dzH1H2] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb104_fixH1] addps xmm1, [esp + nb104_fiyH1] addps xmm2, [esp + nb104_fizH1] movaps [esp + nb104_fjxH2], xmm3 movaps [esp + nb104_fjyH2], xmm4 movaps [esp + nb104_fjzH2], xmm5 movaps [esp + nb104_fixH1], xmm0 movaps [esp + nb104_fiyH1], xmm1 movaps [esp + nb104_fizH1], xmm2 ;# H1-M interaction movaps xmm0, [esp + nb104_rinvH1M] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb104_qqMH] mulps xmm0, xmm1 ;# fs H1-M addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps xmm2, xmm0 xorps xmm3, xmm3 movaps xmm4, xmm3 movaps xmm5, xmm3 mulps xmm0, [esp + nb104_dxH1M] mulps xmm1, [esp + nb104_dyH1M] mulps xmm2, [esp + nb104_dzH1M] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb104_fixH1] addps xmm1, [esp + nb104_fiyH1] addps xmm2, [esp + nb104_fizH1] movaps [esp + nb104_fjxM], xmm3 movaps [esp + nb104_fjyM], xmm4 movaps [esp + nb104_fjzM], xmm5 movaps [esp + nb104_fixH1], xmm0 movaps [esp + nb104_fiyH1], xmm1 movaps [esp + nb104_fizH1], xmm2 ;# H2-H1 interaction movaps xmm0, [esp + nb104_rinvH2H1] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb104_qqHH] mulps xmm0, xmm1 ;# fs H2-H1 addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps xmm2, xmm0 movaps xmm3, [esp + nb104_fjxH1] movaps xmm4, [esp + nb104_fjyH1] movaps xmm5, [esp + nb104_fjzH1] mulps xmm0, [esp + nb104_dxH2H1] mulps xmm1, [esp + nb104_dyH2H1] mulps xmm2, [esp + nb104_dzH2H1] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb104_fixH2] addps xmm1, [esp + nb104_fiyH2] addps xmm2, [esp + nb104_fizH2] movaps [esp + nb104_fjxH1], xmm3 movaps [esp + nb104_fjyH1], xmm4 movaps [esp + nb104_fjzH1], xmm5 movaps [esp + nb104_fixH2], xmm0 movaps [esp + nb104_fiyH2], xmm1 movaps [esp + nb104_fizH2], xmm2 ;# H2-H2 interaction movaps xmm0, [esp + nb104_rinvH2H2] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb104_qqHH] mulps xmm0, xmm1 ;# fsH2H2 addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps xmm2, xmm0 movaps xmm3, [esp + nb104_fjxH2] movaps xmm4, [esp + nb104_fjyH2] movaps xmm5, [esp + nb104_fjzH2] mulps xmm0, [esp + nb104_dxH2H2] mulps xmm1, [esp + nb104_dyH2H2] mulps xmm2, [esp + nb104_dzH2H2] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb104_fixH2] addps xmm1, [esp + nb104_fiyH2] addps xmm2, [esp + nb104_fizH2] movaps [esp + nb104_fjxH2], xmm3 movaps [esp + nb104_fjyH2], xmm4 movaps [esp + nb104_fjzH2], xmm5 movaps [esp + nb104_fixH2], xmm0 movaps [esp + nb104_fiyH2], xmm1 movaps [esp + nb104_fizH2], xmm2 ;# H2-M interaction movaps xmm0, [esp + nb104_rinvH2M] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb104_qqMH] mulps xmm0, xmm1 ;# fs H2-M addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps xmm2, xmm0 movaps xmm3, [esp + nb104_fjxM] movaps xmm4, [esp + nb104_fjyM] movaps xmm5, [esp + nb104_fjzM] mulps xmm0, [esp + nb104_dxH2M] mulps xmm1, [esp + nb104_dyH2M] mulps xmm2, [esp + nb104_dzH2M] subps xmm3, xmm0 subps xmm4, xmm1 subps xmm5, xmm2 addps xmm0, [esp + nb104_fixH2] addps xmm1, [esp + nb104_fiyH2] addps xmm2, [esp + nb104_fizH2] movaps [esp + nb104_fjxM], xmm3 movaps [esp + nb104_fjyM], xmm4 movaps [esp + nb104_fjzM], xmm5 movaps [esp + nb104_fixH2], xmm0 movaps [esp + nb104_fiyH2], xmm1 movaps [esp + nb104_fizH2], xmm2 ;# M-H1 interaction movaps xmm0, [esp + nb104_rinvMH1] movaps xmm1, xmm0 mulps xmm0, xmm0 mulps xmm1, [esp + nb104_qqMH] mulps xmm0, xmm1 ;# fs M-H1 addps xmm7, xmm1 ;# add to local vctot movaps xmm1, xmm0 movaps xmm2, xmm0 movaps xmm3, [esp + nb104_fjxH1] movaps xmm4, [esp + nb104_fjyH1] movaps xmm5, [esp + nb104_fjzH1] mulps xmm0, [esp + nb104_dxMH1] mulps xmm1, [esp + nb104_dyMH1]
⌨️ 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?