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
mulps nb204_rsqH2H1(%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 ## fsOH2 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb204_fjxH1(%esp),%xmm3 movaps nb204_fjyH1(%esp),%xmm4 movaps nb204_fjzH1(%esp),%xmm5 mulps nb204_dxH2H1(%esp),%xmm0 mulps nb204_dyH2H1(%esp),%xmm1 mulps nb204_dzH2H1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb204_fixH2(%esp),%xmm0 addps nb204_fiyH2(%esp),%xmm1 addps nb204_fizH2(%esp),%xmm2 movaps %xmm3,nb204_fjxH1(%esp) movaps %xmm4,nb204_fjyH1(%esp) movaps %xmm5,nb204_fjzH1(%esp) movaps %xmm0,nb204_fixH2(%esp) movaps %xmm1,nb204_fiyH2(%esp) movaps %xmm2,nb204_fizH2(%esp) ## H2-H2 interaction movaps nb204_rinvH2H2(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb204_krf(%esp),%xmm5 movaps %xmm0,%xmm1 mulps nb204_rsqH2H2(%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 ## fsOH2 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb204_fjxH2(%esp),%xmm3 movaps nb204_fjyH2(%esp),%xmm4 movaps nb204_fjzH2(%esp),%xmm5 mulps nb204_dxH2H2(%esp),%xmm0 mulps nb204_dyH2H2(%esp),%xmm1 mulps nb204_dzH2H2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb204_fixH2(%esp),%xmm0 addps nb204_fiyH2(%esp),%xmm1 addps nb204_fizH2(%esp),%xmm2 movaps %xmm3,nb204_fjxH2(%esp) movaps %xmm4,nb204_fjyH2(%esp) movaps %xmm5,nb204_fjzH2(%esp) movaps %xmm0,nb204_fixH2(%esp) movaps %xmm1,nb204_fiyH2(%esp) movaps %xmm2,nb204_fizH2(%esp) ## H2-M interaction movaps nb204_rinvH2M(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb204_krf(%esp),%xmm5 movaps %xmm0,%xmm1 mulps nb204_rsqH2M(%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 movaps nb204_fjxM(%esp),%xmm3 movaps nb204_fjyM(%esp),%xmm4 movaps nb204_fjzM(%esp),%xmm5 mulps nb204_dxH2M(%esp),%xmm0 mulps nb204_dyH2M(%esp),%xmm1 mulps nb204_dzH2M(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb204_fixH2(%esp),%xmm0 addps nb204_fiyH2(%esp),%xmm1 addps nb204_fizH2(%esp),%xmm2 movaps %xmm3,nb204_fjxM(%esp) movaps %xmm4,nb204_fjyM(%esp) movaps %xmm5,nb204_fjzM(%esp) movaps %xmm0,nb204_fixH2(%esp) movaps %xmm1,nb204_fiyH2(%esp) movaps %xmm2,nb204_fizH2(%esp) ## M-H1 interaction movaps nb204_rinvMH1(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb204_krf(%esp),%xmm5 movaps %xmm0,%xmm1 mulps nb204_rsqMH1(%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 movaps nb204_fjxH1(%esp),%xmm3 movaps nb204_fjyH1(%esp),%xmm4 movaps nb204_fjzH1(%esp),%xmm5 mulps nb204_dxMH1(%esp),%xmm0 mulps nb204_dyMH1(%esp),%xmm1 mulps nb204_dzMH1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb204_fixM(%esp),%xmm0 addps nb204_fiyM(%esp),%xmm1 addps nb204_fizM(%esp),%xmm2 movaps %xmm3,nb204_fjxH1(%esp) movaps %xmm4,nb204_fjyH1(%esp) movaps %xmm5,nb204_fjzH1(%esp) movaps %xmm0,nb204_fixM(%esp) movaps %xmm1,nb204_fiyM(%esp) movaps %xmm2,nb204_fizM(%esp) ## M-H2 interaction movaps nb204_rinvMH2(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb204_krf(%esp),%xmm5 movaps %xmm0,%xmm1 mulps nb204_rsqMH2(%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 movaps nb204_fjxH2(%esp),%xmm3 movaps nb204_fjyH2(%esp),%xmm4 movaps nb204_fjzH2(%esp),%xmm5 mulps nb204_dxMH2(%esp),%xmm0 mulps nb204_dyMH2(%esp),%xmm1 mulps nb204_dzMH2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb204_fixM(%esp),%xmm0 addps nb204_fiyM(%esp),%xmm1 addps nb204_fizM(%esp),%xmm2 movaps %xmm3,nb204_fjxH2(%esp) movaps %xmm4,nb204_fjyH2(%esp) movaps %xmm5,nb204_fjzH2(%esp) movaps %xmm0,nb204_fixM(%esp) movaps %xmm1,nb204_fiyM(%esp) movaps %xmm2,nb204_fizM(%esp) ## M-M interaction movaps nb204_rinvMM(%esp),%xmm0 movaps %xmm0,%xmm7 ## xmm7=rinv movaps nb204_krf(%esp),%xmm5 movaps %xmm0,%xmm1 mulps nb204_rsqMM(%esp),%xmm5 ## xmm5=krsq movaps %xmm5,%xmm4 addps %xmm7,%xmm4 ## xmm4=r inv+ krsq subps nb204_crf(%esp),%xmm4 mulps %xmm0,%xmm0 mulps nb204_qqMM(%esp),%xmm4 ## xmm4=voul=qq*(rinv+ krsq-crf) mulps nb204_two(%esp),%xmm5 subps %xmm5,%xmm7 ## xmm7=rinv-2*krsq mulps nb204_qqMM(%esp),%xmm7 ## xmm7 = coul part of fscal addps %xmm4,%xmm6 ## add to local vctot mulps %xmm7,%xmm0 ## fsOH2 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps %xmm0,%xmm1 movaps %xmm6,nb204_vctot(%esp) movaps %xmm0,%xmm2 movaps nb204_fjxM(%esp),%xmm3 movaps nb204_fjyM(%esp),%xmm4 movaps nb204_fjzM(%esp),%xmm5 mulps nb204_dxMM(%esp),%xmm0 mulps nb204_dyMM(%esp),%xmm1 mulps nb204_dzMM(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb204_fixM(%esp),%xmm0 addps nb204_fiyM(%esp),%xmm1 addps nb204_fizM(%esp),%xmm2 movaps %xmm3,nb204_fjxM(%esp) movaps %xmm4,nb204_fjyM(%esp) movaps %xmm5,nb204_fjzM(%esp) movaps %xmm0,nb204_fixM(%esp) movaps %xmm1,nb204_fiyM(%esp) movaps %xmm2,nb204_fizM(%esp) movl nb204_faction(%ebp),%edi ## Did all interactions - now update j forces ## At this stage forces are still on the stack, in positions: ## fjxH1, fjyH1, fjzH1, ... , fjzM. ## Each position is a quadruplet of forces for the four ## corresponding j waters, so we need to transpose them before ## adding to the memory positions. ## ## This _used_ to be a simple transpose, but the resulting high number ## of unaligned 128-bit load/stores might trigger a possible hardware ## bug on Athlon and Opteron chips, so I have worked around it ## to use 64-bit load/stores instead. The performance hit should be ## very modest, since the 128-bit unaligned memory instructions were ## slow anyway. ## 4 j waters with three atoms each - first do 1st Hydrogen X & Y forces for 4 j particles movaps nb204_fjxH1(%esp),%xmm0 ## xmm0= fjxH1a fjxH1b fjxH1c fjxH1d movaps nb204_fjyH1(%esp),%xmm2 ## xmm1= fjyH1a fjyH1b fjyH1c fjyH1d movlps 12(%edi,%eax,4),%xmm3 movlps 12(%edi,%ecx,4),%xmm4 movaps %xmm0,%xmm1 unpcklps %xmm2,%xmm0 ## xmm0= fjxH1a fjyH1a fjxH1b fjyH1b unpckhps %xmm2,%xmm1 ## xmm1= fjxH1c fjyH1c fjxH1d fjyH1d movhps 12(%edi,%ebx,4),%xmm3 movhps 12(%edi,%edx,4),%xmm4 addps %xmm0,%xmm3 addps %xmm1,%xmm4 movlps %xmm3,12(%edi,%eax,4) movlps %xmm4,12(%edi,%ecx,4) movhps %xmm3,12(%edi,%ebx,4) movhps %xmm4,12(%edi,%edx,4) ## constant 1st Hydrogen Z & 2nd hydrogen X forces for 4 j particles movaps nb204_fjzH1(%esp),%xmm0 ## xmm0= fjzH1a fjzH1b fjzH1c fjzH1d movaps nb204_fjxH2(%esp),%xmm2 ## xmm1= fjxH2a fjxH2b fjxH2c fjxH2d movlps 20(%edi,%eax,4),%xmm3 movlps 20(%edi,%ecx,4),%xmm4 movaps %xmm0,%xmm1 unpcklps %xmm2,%xmm0 ## xmm0= fjzH1a fjxH2a fjzH1b fjxH2b unpckhps %xmm2,%xmm1 ## xmm1= fjzH1c fjxH2c fjzH1d fjxH2d movhps 20(%edi,%ebx,4),%xmm3 movhps 20(%edi,%edx,4),%xmm4 addps %xmm0,%xmm3 addps %xmm1,%xmm4 movlps %xmm3,20(%edi,%eax,4) movlps %xmm4,20(%edi,%ecx,4) movhps %xmm3,20(%edi,%ebx,4) movhps %xmm4,20(%edi,%edx,4) ## constant 2nd hydrogen Y & Z forces for 4 j particles movaps nb204_fjyH2(%esp),%xmm0 ## xmm0= fjyH2a fjyH2b fjyH2c fjyH2d movaps nb204_fjzH2(%esp),%xmm2 ## xmm1= fjzH2a fjzH2b fjzH2c fjzH2d movlps 28(%edi,%eax,4),%xmm3 movlps 28(%edi,%ecx,4),%xmm4 movaps %xmm0,%xmm1 unpcklps %xmm2,%xmm0 ## xmm0= fjyH2a fjzH2a fjyH2b fjzH2b unpckhps %xmm2,%xmm1 ## xmm1= fjyH2c fjzH2c fjyH2d fjzH2d movhps 28(%edi,%ebx,4),%xmm3 movhps 28(%edi,%edx,4),%xmm4 addps %xmm0,%xmm3 addps %xmm1,%xmm4 movlps %xmm3,28(%edi,%eax,4) movlps %xmm4,28(%edi,%ecx,4) movhps %xmm3,28(%edi,%ebx,4) movhps %xmm4,28(%edi,%edx,4) ## Dummy (M) X & Y forces for 4 j particles movaps nb204_fjxM(%esp),%xmm0 ## xmm0= fjxMa fjxMb fjxMc fjxMd movaps nb204_fjyM(%esp),%xmm2 ## xmm1= fjyMa fjyMb fjyMc fjyMd movlps 36(%edi,%eax,4),%xmm3 movlps 36(%edi,%ecx,4),%xmm4 movaps %xmm0,%xmm1 unpcklps %xmm2,%xmm0 ## xmm0= fjxMa fjyMa fjxMb fjyMb unpckhps %xmm2,%xmm1 ## xmm1= fjxMc fjyMc fjxMd fjyMd movhps 36(%edi,%ebx,4),%xmm3 movhps 36(%edi,%edx,4),%xmm4 addps %xmm0,%xmm3 addps %xmm1,%xmm4 movlps %xmm3,36(%edi,%eax,4) movlps %xmm4,36(%edi,%ecx,4) movhps %xmm3,36(%edi,%ebx,4) movhps %xmm4,36(%edi,%edx,4) ## Dummy (M) Z forces for 4 j particles ## Just load the four Z coords into one reg. each movss 44(%edi,%eax,4),%xmm4 movss 44(%edi,%ebx,4),%xmm5 movss 44(%edi,%ecx,4),%xmm6 movss 44(%edi,%edx,4),%xmm7 ## add what we have on the stack addss nb204_fjzM(%esp),%xmm4 addss nb204_fjzMb(%esp),%xmm5 addss nb204_fjzMc(%esp),%xmm6 addss nb204_fjzMd(%esp),%xmm7 ## store back movss %xmm4,44(%edi,%eax,4) movss %xmm5,44(%edi,%ebx,4) movss %xmm6,44(%edi,%ecx,4) movss %xmm7,44(%edi,%edx,4) ## should we do one more iteration? subl $4,nb204_innerk(%esp) jl _nb_kernel204_ia32_sse.nb204_single_check jmp _nb_kernel204_ia32_sse.nb204_unroll_loop_nb_kernel204_ia32_sse.nb204_single_check: addl $4,nb204_innerk(%esp) jnz _nb_kernel204_ia32_sse.nb204_single_loop jmp _nb_kernel204_ia32_sse.nb204_updateouterdata_nb_kernel204_ia32_sse.nb204_single_loop: movl nb204_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb204_innerjjnr(%esp) movl nb204_pos(%ebp),%esi leal (%eax,%eax,2),%eax ## fetch j coordinates xorps %xmm3,%xmm3 xorps %xmm4,%xmm4 xorps %xmm5,%xmm5 movss 36(%esi,%eax,4),%xmm3 ## jxM - - - movss 40(%esi,%eax,4),%xmm4 ## jyM - - - movss 44(%esi,%eax,4),%xmm5 ## jzM - - - movlps 12(%esi,%eax,4),%xmm6 ## xmm6 = jxH1 jyH1 - - movss 20(%esi,%eax,4),%xmm7 ## xmm7 = jzH1 - - - movhps 24(%esi,%eax,4),%xmm6 ## xmm6 = jxH1 jyH1 jxH2 jyH2 movss 32(%esi,%eax,4),%xmm2 ## xmm2 = jzH2 - - - ## have all coords, time for some shuffling. shufps $216,%xmm6,%xmm6 ## constant 11011000 ;# xmm6 = jxH1 jxH2 jyH1 jyH2 unpcklps %xmm2,%xmm7 ## xmm7 = jzH1 jzH2 - - movaps nb204_ixM(%esp),%xmm0 movaps nb204_iyM(%esp),%xmm1 movaps nb204_izM(%esp),%xmm2 movlhps %xmm6,%xmm3 ## xmm3 = jxM 0 jxH1 jxH2 shufps $228,%xmm6,%xmm4 ## constant 11100100 ;# xmm4 = jyM 0 jyH1 jyH2 shufps $68,%xmm7,%xmm5 ## constant 01000100 ;# xmm5 = jzM 0 jzH1 jzH2 ## store all j coordinates in jM movaps %xmm3,nb204_jxM(%esp) movaps %xmm4,nb204_jyM(%esp) movaps %xmm5,nb204_jzM(%esp) subps %xmm3,%xmm0 subps %xmm4,%xmm1 subps %xmm5,%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 ## have rsq in xmm0 movaps %xmm0,%xmm6 ## do invsqrt rsqrtps %xmm0,%xmm1 mulps nb204_krf(%esp),%xmm6 ## xmm6=krsq movaps %xmm1,%xmm2 movaps %xmm6,%xmm7 ## xmm7=krsq ⌨️ 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?