nb_kernel204_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,959 行 · 第 1/5 页
S
1,959 行
unpckhps %xmm5,%xmm2 unpckhps %xmm6,%xmm3 movaps %xmm4,%xmm5 movaps %xmm0,%xmm6 unpcklps %xmm7,%xmm4 unpckhps %xmm7,%xmm5 movaps %xmm1,%xmm7 movlhps %xmm2,%xmm0 movaps %xmm0,nb204_jxH1(%esp) movhlps %xmm6,%xmm2 movaps %xmm2,nb204_jyH1(%esp) movlhps %xmm3,%xmm1 movaps %xmm1,nb204_jxH2(%esp) movhlps %xmm7,%xmm3 movaps %xmm4,%xmm6 movaps %xmm3,nb204_jyH2(%esp) movlhps %xmm5,%xmm4 movaps %xmm4,nb204_jxM(%esp) movhlps %xmm6,%xmm5 movaps %xmm5,nb204_jyM(%esp) movss 20(%esi,%eax,4),%xmm0 movss 32(%esi,%eax,4),%xmm1 movss 44(%esi,%eax,4),%xmm2 movss 20(%esi,%ecx,4),%xmm3 movss 32(%esi,%ecx,4),%xmm4 movss 44(%esi,%ecx,4),%xmm5 movhps 16(%esi,%ebx,4),%xmm0 movhps 28(%esi,%ebx,4),%xmm1 movhps 40(%esi,%ebx,4),%xmm2 movhps 16(%esi,%edx,4),%xmm3 movhps 28(%esi,%edx,4),%xmm4 movhps 40(%esi,%edx,4),%xmm5 shufps $204,%xmm3,%xmm0 ## constant 11001100 shufps $204,%xmm4,%xmm1 ## constant 11001100 shufps $204,%xmm5,%xmm2 ## constant 11001100 movaps %xmm0,nb204_jzH1(%esp) movaps %xmm1,nb204_jzH2(%esp) movaps %xmm2,nb204_jzM(%esp) movaps nb204_ixH1(%esp),%xmm0 movaps nb204_iyH1(%esp),%xmm1 movaps nb204_izH1(%esp),%xmm2 movaps nb204_ixH1(%esp),%xmm3 movaps nb204_iyH1(%esp),%xmm4 movaps nb204_izH1(%esp),%xmm5 subps nb204_jxH1(%esp),%xmm0 subps nb204_jyH1(%esp),%xmm1 subps nb204_jzH1(%esp),%xmm2 subps nb204_jxH2(%esp),%xmm3 subps nb204_jyH2(%esp),%xmm4 subps nb204_jzH2(%esp),%xmm5 movaps %xmm0,nb204_dxH1H1(%esp) movaps %xmm1,nb204_dyH1H1(%esp) movaps %xmm2,nb204_dzH1H1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb204_dxH1H2(%esp) movaps %xmm4,nb204_dyH1H2(%esp) movaps %xmm5,nb204_dzH1H2(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb204_rsqH1H1(%esp) movaps %xmm3,nb204_rsqH1H2(%esp) movaps nb204_ixH1(%esp),%xmm0 movaps nb204_iyH1(%esp),%xmm1 movaps nb204_izH1(%esp),%xmm2 movaps nb204_ixH2(%esp),%xmm3 movaps nb204_iyH2(%esp),%xmm4 movaps nb204_izH2(%esp),%xmm5 subps nb204_jxM(%esp),%xmm0 subps nb204_jyM(%esp),%xmm1 subps nb204_jzM(%esp),%xmm2 subps nb204_jxH1(%esp),%xmm3 subps nb204_jyH1(%esp),%xmm4 subps nb204_jzH1(%esp),%xmm5 movaps %xmm0,nb204_dxH1M(%esp) movaps %xmm1,nb204_dyH1M(%esp) movaps %xmm2,nb204_dzH1M(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb204_dxH2H1(%esp) movaps %xmm4,nb204_dyH2H1(%esp) movaps %xmm5,nb204_dzH2H1(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb204_rsqH1M(%esp) movaps %xmm3,nb204_rsqH2H1(%esp) movaps nb204_ixH2(%esp),%xmm0 movaps nb204_iyH2(%esp),%xmm1 movaps nb204_izH2(%esp),%xmm2 movaps nb204_ixH2(%esp),%xmm3 movaps nb204_iyH2(%esp),%xmm4 movaps nb204_izH2(%esp),%xmm5 subps nb204_jxH2(%esp),%xmm0 subps nb204_jyH2(%esp),%xmm1 subps nb204_jzH2(%esp),%xmm2 subps nb204_jxM(%esp),%xmm3 subps nb204_jyM(%esp),%xmm4 subps nb204_jzM(%esp),%xmm5 movaps %xmm0,nb204_dxH2H2(%esp) movaps %xmm1,nb204_dyH2H2(%esp) movaps %xmm2,nb204_dzH2H2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb204_dxH2M(%esp) movaps %xmm4,nb204_dyH2M(%esp) movaps %xmm5,nb204_dzH2M(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb204_rsqH2H2(%esp) movaps %xmm3,nb204_rsqH2M(%esp) movaps nb204_ixM(%esp),%xmm0 movaps nb204_iyM(%esp),%xmm1 movaps nb204_izM(%esp),%xmm2 movaps nb204_ixM(%esp),%xmm3 movaps nb204_iyM(%esp),%xmm4 movaps nb204_izM(%esp),%xmm5 subps nb204_jxH1(%esp),%xmm0 subps nb204_jyH1(%esp),%xmm1 subps nb204_jzH1(%esp),%xmm2 subps nb204_jxH2(%esp),%xmm3 subps nb204_jyH2(%esp),%xmm4 subps nb204_jzH2(%esp),%xmm5 movaps %xmm0,nb204_dxMH1(%esp) movaps %xmm1,nb204_dyMH1(%esp) movaps %xmm2,nb204_dzMH1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb204_dxMH2(%esp) movaps %xmm4,nb204_dyMH2(%esp) movaps %xmm5,nb204_dzMH2(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm3,%xmm4 addps %xmm5,%xmm4 movaps %xmm0,nb204_rsqMH1(%esp) movaps %xmm4,nb204_rsqMH2(%esp) movaps nb204_ixM(%esp),%xmm0 movaps nb204_iyM(%esp),%xmm1 movaps nb204_izM(%esp),%xmm2 subps nb204_jxM(%esp),%xmm0 subps nb204_jyM(%esp),%xmm1 subps nb204_jzM(%esp),%xmm2 movaps %xmm0,nb204_dxMM(%esp) movaps %xmm1,nb204_dyMM(%esp) movaps %xmm2,nb204_dzMM(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 movaps %xmm0,nb204_rsqMM(%esp) ## start doing invsqrt use rsq values in xmm0, xmm4 rsqrtps %xmm0,%xmm1 rsqrtps %xmm4,%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb204_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps %xmm0,%xmm1 mulps %xmm4,%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb204_half(%esp),%xmm3 ## rinvH2H2 mulps nb204_half(%esp),%xmm7 ## rinvH2H1 movaps %xmm3,nb204_rinvMM(%esp) movaps %xmm7,nb204_rinvMH2(%esp) rsqrtps nb204_rsqH1H1(%esp),%xmm1 rsqrtps nb204_rsqH1H2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb204_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb204_rsqH1H1(%esp),%xmm1 mulps nb204_rsqH1H2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb204_half(%esp),%xmm3 mulps nb204_half(%esp),%xmm7 movaps %xmm3,nb204_rinvH1H1(%esp) movaps %xmm7,nb204_rinvH1H2(%esp) rsqrtps nb204_rsqH1M(%esp),%xmm1 rsqrtps nb204_rsqH2H1(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb204_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb204_rsqH1M(%esp),%xmm1 mulps nb204_rsqH2H1(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb204_half(%esp),%xmm3 mulps nb204_half(%esp),%xmm7 movaps %xmm3,nb204_rinvH1M(%esp) movaps %xmm7,nb204_rinvH2H1(%esp) rsqrtps nb204_rsqH2H2(%esp),%xmm1 rsqrtps nb204_rsqH2M(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb204_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb204_rsqH2H2(%esp),%xmm1 mulps nb204_rsqH2M(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb204_half(%esp),%xmm3 mulps nb204_half(%esp),%xmm7 movaps %xmm3,nb204_rinvH2H2(%esp) movaps %xmm7,nb204_rinvH2M(%esp) rsqrtps nb204_rsqMH1(%esp),%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb204_three(%esp),%xmm3 mulps nb204_rsqMH1(%esp),%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb204_half(%esp),%xmm3 movaps %xmm3,nb204_rinvMH1(%esp) ## start with H1-H1 interaction movaps nb204_rinvH1H1(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb204_krf(%esp),%xmm5 mulps %xmm0,%xmm0 ## xmm0=rinvsq mulps nb204_rsqH1H1(%esp),%xmm5 ## xmm5=krsq movaps %xmm5,%xmm6 addps %xmm7,%xmm6 ## xmm6=rinv+ krsq subps nb204_crf(%esp),%xmm6 mulps nb204_qqHH(%esp),%xmm6 ## xmm6=voul=qq*(rinv+ krsq-crf) mulps nb204_two(%esp),%xmm5 subps %xmm5,%xmm7 ## xmm7=rinv-2*krsq mulps nb204_qqHH(%esp),%xmm7 ## xmm7 = coul part of fscal addps nb204_vctot(%esp),%xmm6 ## local vctot summation variable mulps %xmm7,%xmm0 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb204_dxH1H1(%esp),%xmm0 mulps nb204_dyH1H1(%esp),%xmm1 mulps nb204_dzH1H1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb204_fixH1(%esp),%xmm0 addps nb204_fiyH1(%esp),%xmm1 addps nb204_fizH1(%esp),%xmm2 movaps %xmm3,nb204_fjxH1(%esp) movaps %xmm4,nb204_fjyH1(%esp) movaps %xmm5,nb204_fjzH1(%esp) movaps %xmm0,nb204_fixH1(%esp) movaps %xmm1,nb204_fiyH1(%esp) movaps %xmm2,nb204_fizH1(%esp) ## H1-H2 interaction movaps nb204_rinvH1H2(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb204_krf(%esp),%xmm5 movaps %xmm0,%xmm1 mulps nb204_rsqH1H2(%esp),%xmm5 ## xmm5=krsq movaps %xmm5,%xmm4 addps %xmm7,%xmm4 ## xmm4=r inv+ krsq subps nb204_crf(%esp),%xmm4 mulps %xmm0,%xmm0 mulps nb204_qqHH(%esp),%xmm4 ## xmm4=voul=qq*(rinv+ krsq-crf) mulps nb204_two(%esp),%xmm5 subps %xmm5,%xmm7 ## xmm7=rinv-2*krsq mulps nb204_qqHH(%esp),%xmm7 ## xmm7 = coul part of fscal addps %xmm4,%xmm6 ## add to local vctot mulps %xmm7,%xmm0 ## fsOH1 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb204_dxH1H2(%esp),%xmm0 mulps nb204_dyH1H2(%esp),%xmm1 mulps nb204_dzH1H2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb204_fixH1(%esp),%xmm0 addps nb204_fiyH1(%esp),%xmm1 addps nb204_fizH1(%esp),%xmm2 movaps %xmm3,nb204_fjxH2(%esp) movaps %xmm4,nb204_fjyH2(%esp) movaps %xmm5,nb204_fjzH2(%esp) movaps %xmm0,nb204_fixH1(%esp) movaps %xmm1,nb204_fiyH1(%esp) movaps %xmm2,nb204_fizH1(%esp) ## H1-M interaction movaps nb204_rinvH1M(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=Rinv movaps nb204_krf(%esp),%xmm5 movaps %xmm0,%xmm1 mulps nb204_rsqH1M(%esp),%xmm5 ## xmm5=krsq movaps %xmm5,%xmm4 addps %xmm7,%xmm4 ## xmm4=r inv+ krsq subps nb204_crf(%esp),%xmm4 mulps %xmm0,%xmm0 mulps nb204_qqMH(%esp),%xmm4 ## xmm4=voul=qq*(rinv+ krsq-crf) mulps nb204_two(%esp),%xmm5 subps %xmm5,%xmm7 ## xmm7=rinv-2*krsq mulps nb204_qqMH(%esp),%xmm7 ## xmm7 = coul part of fscal addps %xmm4,%xmm6 ## add to local vctot mulps %xmm7,%xmm0 ## fsOH2 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb204_dxH1M(%esp),%xmm0 mulps nb204_dyH1M(%esp),%xmm1 mulps nb204_dzH1M(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb204_fixH1(%esp),%xmm0 addps nb204_fiyH1(%esp),%xmm1 addps nb204_fizH1(%esp),%xmm2 movaps %xmm3,nb204_fjxM(%esp) movaps %xmm4,nb204_fjyM(%esp) movaps %xmm5,nb204_fjzM(%esp) movaps %xmm0,nb204_fixH1(%esp) movaps %xmm1,nb204_fiyH1(%esp) movaps %xmm2,nb204_fizH1(%esp) ## H2-H1 interaction movaps nb204_rinvH2H1(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb204_krf(%esp),%xmm5 movaps %xmm0,%xmm1
⌨️ 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?