nb_kernel114_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,017 行 · 第 1/5 页
S
2,017 行
movaps nb114_rinvH1H2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb114_qqHH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs H1-H2 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb114_dxH1H2(%esp),%xmm0 mulps nb114_dyH1H2(%esp),%xmm1 mulps nb114_dzH1H2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb114_fixH1(%esp),%xmm0 addps nb114_fiyH1(%esp),%xmm1 addps nb114_fizH1(%esp),%xmm2 movaps %xmm3,nb114_fjxH2(%esp) movaps %xmm4,nb114_fjyH2(%esp) movaps %xmm5,nb114_fjzH2(%esp) movaps %xmm0,nb114_fixH1(%esp) movaps %xmm1,nb114_fiyH1(%esp) movaps %xmm2,nb114_fizH1(%esp) ## H1-M interaction movaps nb114_rinvH1M(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb114_qqMH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs H1-M addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb114_dxH1M(%esp),%xmm0 mulps nb114_dyH1M(%esp),%xmm1 mulps nb114_dzH1M(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb114_fixH1(%esp),%xmm0 addps nb114_fiyH1(%esp),%xmm1 addps nb114_fizH1(%esp),%xmm2 movaps %xmm3,nb114_fjxM(%esp) movaps %xmm4,nb114_fjyM(%esp) movaps %xmm5,nb114_fjzM(%esp) movaps %xmm0,nb114_fixH1(%esp) movaps %xmm1,nb114_fiyH1(%esp) movaps %xmm2,nb114_fizH1(%esp) ## H2-H1 interaction movaps nb114_rinvH2H1(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb114_qqHH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs H2-H1 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb114_fjxH1(%esp),%xmm3 movaps nb114_fjyH1(%esp),%xmm4 movaps nb114_fjzH1(%esp),%xmm5 mulps nb114_dxH2H1(%esp),%xmm0 mulps nb114_dyH2H1(%esp),%xmm1 mulps nb114_dzH2H1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb114_fixH2(%esp),%xmm0 addps nb114_fiyH2(%esp),%xmm1 addps nb114_fizH2(%esp),%xmm2 movaps %xmm3,nb114_fjxH1(%esp) movaps %xmm4,nb114_fjyH1(%esp) movaps %xmm5,nb114_fjzH1(%esp) movaps %xmm0,nb114_fixH2(%esp) movaps %xmm1,nb114_fiyH2(%esp) movaps %xmm2,nb114_fizH2(%esp) ## H2-H2 interaction movaps nb114_rinvH2H2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb114_qqHH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fsH2H2 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb114_fjxH2(%esp),%xmm3 movaps nb114_fjyH2(%esp),%xmm4 movaps nb114_fjzH2(%esp),%xmm5 mulps nb114_dxH2H2(%esp),%xmm0 mulps nb114_dyH2H2(%esp),%xmm1 mulps nb114_dzH2H2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb114_fixH2(%esp),%xmm0 addps nb114_fiyH2(%esp),%xmm1 addps nb114_fizH2(%esp),%xmm2 movaps %xmm3,nb114_fjxH2(%esp) movaps %xmm4,nb114_fjyH2(%esp) movaps %xmm5,nb114_fjzH2(%esp) movaps %xmm0,nb114_fixH2(%esp) movaps %xmm1,nb114_fiyH2(%esp) movaps %xmm2,nb114_fizH2(%esp) ## H2-M interaction movaps nb114_rinvH2M(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb114_qqMH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs H2-M addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb114_fjxM(%esp),%xmm3 movaps nb114_fjyM(%esp),%xmm4 movaps nb114_fjzM(%esp),%xmm5 mulps nb114_dxH2M(%esp),%xmm0 mulps nb114_dyH2M(%esp),%xmm1 mulps nb114_dzH2M(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb114_fixH2(%esp),%xmm0 addps nb114_fiyH2(%esp),%xmm1 addps nb114_fizH2(%esp),%xmm2 movaps %xmm3,nb114_fjxM(%esp) movaps %xmm4,nb114_fjyM(%esp) movaps %xmm5,nb114_fjzM(%esp) movaps %xmm0,nb114_fixH2(%esp) movaps %xmm1,nb114_fiyH2(%esp) movaps %xmm2,nb114_fizH2(%esp) ## M-H1 interaction movaps nb114_rinvMH1(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb114_qqMH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs M-H1 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb114_fjxH1(%esp),%xmm3 movaps nb114_fjyH1(%esp),%xmm4 movaps nb114_fjzH1(%esp),%xmm5 mulps nb114_dxMH1(%esp),%xmm0 mulps nb114_dyMH1(%esp),%xmm1 mulps nb114_dzMH1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb114_fixM(%esp),%xmm0 addps nb114_fiyM(%esp),%xmm1 addps nb114_fizM(%esp),%xmm2 movaps %xmm3,nb114_fjxH1(%esp) movaps %xmm4,nb114_fjyH1(%esp) movaps %xmm5,nb114_fjzH1(%esp) movaps %xmm0,nb114_fixM(%esp) movaps %xmm1,nb114_fiyM(%esp) movaps %xmm2,nb114_fizM(%esp) ## M-H2 interaction movaps nb114_rinvMH2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb114_qqMH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs M-H2 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb114_fjxH2(%esp),%xmm3 movaps nb114_fjyH2(%esp),%xmm4 movaps nb114_fjzH2(%esp),%xmm5 mulps nb114_dxMH2(%esp),%xmm0 mulps nb114_dyMH2(%esp),%xmm1 mulps nb114_dzMH2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb114_fixM(%esp),%xmm0 addps nb114_fiyM(%esp),%xmm1 addps nb114_fizM(%esp),%xmm2 movaps %xmm3,nb114_fjxH2(%esp) movaps %xmm4,nb114_fjyH2(%esp) movaps %xmm5,nb114_fjzH2(%esp) movaps %xmm0,nb114_fixM(%esp) movaps %xmm1,nb114_fiyM(%esp) movaps %xmm2,nb114_fizM(%esp) ## M-M interaction movaps nb114_rinvMM(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb114_qqMM(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs M-M addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm7,nb114_vctot(%esp) movaps %xmm0,%xmm2 movaps nb114_fjxM(%esp),%xmm3 movaps nb114_fjyM(%esp),%xmm4 movaps nb114_fjzM(%esp),%xmm5 mulps nb114_dxMM(%esp),%xmm0 mulps nb114_dyMM(%esp),%xmm1 mulps nb114_dzMM(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb114_fixM(%esp),%xmm0 addps nb114_fiyM(%esp),%xmm1 addps nb114_fizM(%esp),%xmm2 movaps %xmm3,nb114_fjxM(%esp) movaps %xmm4,nb114_fjyM(%esp) movaps %xmm5,nb114_fjzM(%esp) movaps %xmm0,nb114_fixM(%esp) movaps %xmm1,nb114_fiyM(%esp) movaps %xmm2,nb114_fizM(%esp) ## Did all interactions - now update j forces movl nb114_faction(%ebp),%edi ## 4 j waters with four atoms each. ## step 1 : transpose fjxO, fjyO, fjzO, fjxH1 movaps nb114_fjxO(%esp),%xmm0 movaps nb114_fjyO(%esp),%xmm1 movaps nb114_fjzO(%esp),%xmm2 movaps nb114_fjxH1(%esp),%xmm3 movaps %xmm0,%xmm4 movaps %xmm1,%xmm5 unpcklps %xmm2,%xmm4 unpcklps %xmm3,%xmm5 unpckhps %xmm2,%xmm0 unpckhps %xmm3,%xmm1 movaps %xmm4,%xmm2 movaps %xmm0,%xmm3 unpcklps %xmm5,%xmm4 unpckhps %xmm5,%xmm2 unpcklps %xmm1,%xmm0 unpckhps %xmm1,%xmm3 ## results are now in xmm4, xmm2, xmm0, xmm3 ## load the corresponding j forces from memory movlps (%edi,%eax,4),%xmm1 movlps (%edi,%ebx,4),%xmm5 movlps (%edi,%ecx,4),%xmm6 movlps (%edi,%edx,4),%xmm7 movhps 8(%edi,%eax,4),%xmm1 movhps 8(%edi,%ebx,4),%xmm5 movhps 8(%edi,%ecx,4),%xmm6 movhps 8(%edi,%edx,4),%xmm7 ## add addps %xmm4,%xmm1 addps %xmm2,%xmm5 addps %xmm0,%xmm6 addps %xmm3,%xmm7 ## store back movlps %xmm1,(%edi,%eax,4) movlps %xmm5,(%edi,%ebx,4) movlps %xmm6,(%edi,%ecx,4) movlps %xmm7,(%edi,%edx,4) movhps %xmm1,8(%edi,%eax,4) movhps %xmm5,8(%edi,%ebx,4) movhps %xmm6,8(%edi,%ecx,4) movhps %xmm7,8(%edi,%edx,4) ## step 2 : transpose fjyH1, fjzH1, fjxH2, fjyH2 movaps nb114_fjyH1(%esp),%xmm0 movaps nb114_fjzH1(%esp),%xmm1 movaps nb114_fjxH2(%esp),%xmm2 movaps nb114_fjyH2(%esp),%xmm3 movaps %xmm0,%xmm4 movaps %xmm1,%xmm5 unpcklps %xmm2,%xmm4 unpcklps %xmm3,%xmm5 unpckhps %xmm2,%xmm0 unpckhps %xmm3,%xmm1 movaps %xmm4,%xmm2 movaps %xmm0,%xmm3 unpcklps %xmm5,%xmm4 unpckhps %xmm5,%xmm2 unpcklps %xmm1,%xmm0 unpckhps %xmm1,%xmm3 ## results are now in xmm4, xmm2, xmm0, xmm3 ## load the corresponding j forces from memory movlps 16(%edi,%eax,4),%xmm1 movlps 16(%edi,%ebx,4),%xmm5 movlps 16(%edi,%ecx,4),%xmm6 movlps 16(%edi,%edx,4),%xmm7 movhps 24(%edi,%eax,4),%xmm1 movhps 24(%edi,%ebx,4),%xmm5 movhps 24(%edi,%ecx,4),%xmm6 movhps 24(%edi,%edx,4),%xmm7 ## add addps %xmm4,%xmm1 addps %xmm2,%xmm5 addps %xmm0,%xmm6 addps %xmm3,%xmm7 ## store back movlps %xmm1,16(%edi,%eax,4) movlps %xmm5,16(%edi,%ebx,4) movlps %xmm6,16(%edi,%ecx,4) movlps %xmm7,16(%edi,%edx,4) movhps %xmm1,24(%edi,%eax,4) movhps %xmm5,24(%edi,%ebx,4) movhps %xmm6,24(%edi,%ecx,4) movhps %xmm7,24(%edi,%edx,4) ## step 3 : transpose fjzH2, fjxM, fjyM, fjzM. xmm4 is scratch movaps nb114_fjzH2(%esp),%xmm0 movaps nb114_fjxM(%esp),%xmm1 movaps nb114_fjyM(%esp),%xmm2 movaps nb114_fjzM(%esp),%xmm3 movaps %xmm0,%xmm4 movaps %xmm1,%xmm5 unpcklps %xmm2,%xmm4 unpcklps %xmm3,%xmm5 unpckhps %xmm2,%xmm0 unpckhps %xmm3,%xmm1 movaps %xmm4,%xmm2 movaps %xmm0,%xmm3 unpcklps %xmm5,%xmm4 unpckhps %xmm5,%xmm2 unpcklps %xmm1,%xmm0 unpckhps %xmm1,%xmm3 ## results are now in xmm0, xmm1, xmm2, xmm3 ## load the corresponding j forces from memory movlps 32(%edi,%eax,4),%xmm1 movlps 32(%edi,%ebx,4),%xmm5 movlps 32(%edi,%ecx,4),%xmm6 movlps 32(%edi,%edx,4),%xmm7 movhps 40(%edi,%eax,4),%xmm1 movhps 40(%edi,%ebx,4),%xmm5 movhps 40(%edi,%ecx,4),%xmm6 movhps 40(%edi,%edx,4),%xmm7 ## add addps %xmm4,%xmm1 addps %xmm2,%xmm5 addps %xmm0,%xmm6 addps %xmm3,%xmm7 ## store back movlps %xmm1,32(%edi,%eax,4) movlps %xmm5,32(%edi,%ebx,4) movlps %xmm6,32(%edi,%ecx,4) movlps %xmm7,32(%edi,%edx,4) movhps %xmm1,40(%edi,%eax,4) movhps %xmm5,40(%edi,%ebx,4) movhps %xmm6,40(%edi,%ecx,4) movhps %xmm7,40(%edi,%edx,4) ## should we do one more iteration? subl $4,nb114_innerk(%esp) jl _nb_kernel114_ia32_sse.nb114_single_check jmp _nb_kernel114_ia32_sse.nb114_unroll_loop_nb_kernel114_ia32_sse.nb114_single_check: addl $4,nb114_innerk(%esp) jnz _nb_kernel114_ia32_sse.nb114_single_loop jmp _nb_kernel114_ia32_sse.nb114_updateouterdata_nb_kernel114_ia32_sse.nb114_single_loop: movl nb114_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb114_innerjjnr(%esp) movl nb114_pos(%ebp),%esi leal (%eax,%eax,2),%eax ## fetch j coordinates movlps (%esi,%eax,4),%xmm3 ## Ox Oy movlps 16(%esi,%eax,4),%xmm4 ## H1y H1z movlps 32(%esi,%eax,4),%xmm5 ## H2z Mx movhps 8(%esi,%eax,4),%xmm3 ## Ox Oy Oz H1x movhps 24(%esi,%eax,4),%xmm4 ## H1y H1z H2x H2y movhps 40(%esi,%eax,4),%xmm5 ## H2z Mx My Mz ## transpose movaps %xmm4,%xmm0 movaps %xmm3,%xmm1 movaps %xmm4,%xmm2 movaps %xmm3,%xmm6 shufps $18,%xmm5,%xmm4 ## (00010010) h2x - Mx - shufps $193,%xmm0,%xmm3 ## (11000001) Oy - H1y - shufps $35,%xmm5,%xmm2 ## (00100011) H2y - My - shufps $18,%xmm0,%xmm1 ## (00010010) Oz - H1z - ## xmm6: Ox - - H1x xmm5: H2z - - Mz shufps $140,%xmm4,%xmm6 ## (10001100) Ox H1x H2x Mx shufps $136,%xmm2,%xmm3 ## (10001000) Oy H1y H2y My shufps $200,%xmm5,%xmm1 ## (11001000) Oz H1z H2z Mz ## store all j coordinates in jO movaps %xmm6,nb114_jxO(%esp) movaps %xmm3,nb114_jyO(%esp) movaps %xmm1,nb114_jzO(%esp) ## do O and M in parallel⌨️ 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?