nb_kernel102_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,986 行 · 第 1/5 页
S
1,986 行
movaps %xmm5,nb102_fjzH1(%esp) movaps %xmm0,nb102_fixH1(%esp) movaps %xmm1,nb102_fiyH1(%esp) movaps %xmm2,nb102_fizH1(%esp) ## H1-H2 interaction movaps nb102_rinvH1H2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb102_qqHH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fsOH2 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb102_fjxH2(%esp),%xmm3 movaps nb102_fjyH2(%esp),%xmm4 movaps nb102_fjzH2(%esp),%xmm5 mulps nb102_dxH1H2(%esp),%xmm0 mulps nb102_dyH1H2(%esp),%xmm1 mulps nb102_dzH1H2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb102_fixH1(%esp),%xmm0 addps nb102_fiyH1(%esp),%xmm1 addps nb102_fizH1(%esp),%xmm2 movaps %xmm3,nb102_fjxH2(%esp) movaps %xmm4,nb102_fjyH2(%esp) movaps %xmm5,nb102_fjzH2(%esp) movaps %xmm0,nb102_fixH1(%esp) movaps %xmm1,nb102_fiyH1(%esp) movaps %xmm2,nb102_fizH1(%esp) ## H2-O interaction movaps nb102_rinvH2O(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb102_qqOH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fsH2O addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb102_fjxO(%esp),%xmm3 movaps nb102_fjyO(%esp),%xmm4 movaps nb102_fjzO(%esp),%xmm5 mulps nb102_dxH2O(%esp),%xmm0 mulps nb102_dyH2O(%esp),%xmm1 mulps nb102_dzH2O(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb102_fixH2(%esp),%xmm0 addps nb102_fiyH2(%esp),%xmm1 addps nb102_fizH2(%esp),%xmm2 movaps %xmm3,nb102_fjxO(%esp) movaps %xmm4,nb102_fjyO(%esp) movaps %xmm5,nb102_fjzO(%esp) movaps %xmm0,nb102_fixH2(%esp) movaps %xmm1,nb102_fiyH2(%esp) movaps %xmm2,nb102_fizH2(%esp) ## H2-H1 interaction movaps nb102_rinvH2H1(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb102_qqHH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fsH2H1 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb102_fjxH1(%esp),%xmm3 movaps nb102_fjyH1(%esp),%xmm4 movaps nb102_fjzH1(%esp),%xmm5 mulps nb102_dxH2H1(%esp),%xmm0 mulps nb102_dyH2H1(%esp),%xmm1 mulps nb102_dzH2H1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb102_fixH2(%esp),%xmm0 addps nb102_fiyH2(%esp),%xmm1 addps nb102_fizH2(%esp),%xmm2 movaps %xmm3,nb102_fjxH1(%esp) movaps %xmm4,nb102_fjyH1(%esp) movaps %xmm5,nb102_fjzH1(%esp) movaps %xmm0,nb102_fixH2(%esp) movaps %xmm1,nb102_fiyH2(%esp) movaps %xmm2,nb102_fizH2(%esp) ## H2-H2 interaction movaps nb102_rinvH2H2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb102_qqHH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fsH2H2 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm7,nb102_vctot(%esp) movaps %xmm0,%xmm2 movaps nb102_fjxH2(%esp),%xmm3 movaps nb102_fjyH2(%esp),%xmm4 movaps nb102_fjzH2(%esp),%xmm5 mulps nb102_dxH2H2(%esp),%xmm0 mulps nb102_dyH2H2(%esp),%xmm1 mulps nb102_dzH2H2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb102_fixH2(%esp),%xmm0 addps nb102_fiyH2(%esp),%xmm1 addps nb102_fizH2(%esp),%xmm2 movaps %xmm3,nb102_fjxH2(%esp) movaps %xmm4,nb102_fjyH2(%esp) movaps %xmm5,nb102_fjzH2(%esp) movaps %xmm0,nb102_fixH2(%esp) movaps %xmm1,nb102_fiyH2(%esp) movaps %xmm2,nb102_fizH2(%esp) movl nb102_faction(%ebp),%edi ## Did all interactions - now update j forces ## At this stage forces are still on the stack, in positions: ## fjxO, fjyO, fjzO, ... , fjzH2. ## 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 Oxygen X & Y forces for 4 j particles movaps nb102_fjxO(%esp),%xmm0 ## xmm0= fjxOa fjxOb fjxOc fjxOd movaps nb102_fjyO(%esp),%xmm2 ## xmm1= fjyOa fjyOb fjyOc fjyOd movlps (%edi,%eax,4),%xmm3 movlps (%edi,%ecx,4),%xmm4 movaps %xmm0,%xmm1 unpcklps %xmm2,%xmm0 ## xmm0= fjxOa fjyOa fjxOb fjyOb unpckhps %xmm2,%xmm1 ## xmm1= fjxOc fjyOc fjxOd fjyOd movhps (%edi,%ebx,4),%xmm3 movhps (%edi,%edx,4),%xmm4 addps %xmm0,%xmm3 addps %xmm1,%xmm4 movlps %xmm3,(%edi,%eax,4) movlps %xmm4,(%edi,%ecx,4) movhps %xmm3,(%edi,%ebx,4) movhps %xmm4,(%edi,%edx,4) ## Oxygen Z & first hydrogen X forces for 4 j particles movaps nb102_fjzO(%esp),%xmm0 ## xmm0= fjzOa fjzOb fjzOc fjzOd movaps nb102_fjxH1(%esp),%xmm2 ## xmm1= fjxH1a fjxH1b fjxH1c fjxH1d movlps 8(%edi,%eax,4),%xmm3 movlps 8(%edi,%ecx,4),%xmm4 movaps %xmm0,%xmm1 unpcklps %xmm2,%xmm0 ## xmm0= fjzOa fjxH1a fjzOb fjxH1b unpckhps %xmm2,%xmm1 ## xmm1= fjzOc fjxH1c fjzOd fjxH1d movhps 8(%edi,%ebx,4),%xmm3 movhps 8(%edi,%edx,4),%xmm4 addps %xmm0,%xmm3 addps %xmm1,%xmm4 movlps %xmm3,8(%edi,%eax,4) movlps %xmm4,8(%edi,%ecx,4) movhps %xmm3,8(%edi,%ebx,4) movhps %xmm4,8(%edi,%edx,4) ## First hydrogen Y & Z forces for 4 j particles movaps nb102_fjyH1(%esp),%xmm0 ## xmm0= fjyH1a fjyH1b fjyH1c fjyH1d movaps nb102_fjzH1(%esp),%xmm2 ## xmm1= fjzH1a fjzH1b fjzH1c fjzH1d movlps 16(%edi,%eax,4),%xmm3 movlps 16(%edi,%ecx,4),%xmm4 movaps %xmm0,%xmm1 unpcklps %xmm2,%xmm0 ## xmm0= fjyH1a fjzH1a fjyH1b fjzH1b unpckhps %xmm2,%xmm1 ## xmm1= fjyH1c fjzH1c fjyH1d fjzH1d movhps 16(%edi,%ebx,4),%xmm3 movhps 16(%edi,%edx,4),%xmm4 addps %xmm0,%xmm3 addps %xmm1,%xmm4 movlps %xmm3,16(%edi,%eax,4) movlps %xmm4,16(%edi,%ecx,4) movhps %xmm3,16(%edi,%ebx,4) movhps %xmm4,16(%edi,%edx,4) ## Second hydrogen X & Y forces for 4 j particles movaps nb102_fjxH2(%esp),%xmm0 ## xmm0= fjxH2a fjxH2b fjxH2c fjxH2d movaps nb102_fjyH2(%esp),%xmm2 ## xmm1= fjyH2a fjyH2b fjyH2c fjyH2d movlps 24(%edi,%eax,4),%xmm3 movlps 24(%edi,%ecx,4),%xmm4 movaps %xmm0,%xmm1 unpcklps %xmm2,%xmm0 ## xmm0= fjxH2a fjyH2a fjxH2b fjyH2b unpckhps %xmm2,%xmm1 ## xmm1= fjxH2c fjyH2c fjxH2d fjyH2d movhps 24(%edi,%ebx,4),%xmm3 movhps 24(%edi,%edx,4),%xmm4 addps %xmm0,%xmm3 addps %xmm1,%xmm4 movlps %xmm3,24(%edi,%eax,4) movlps %xmm4,24(%edi,%ecx,4) movhps %xmm3,24(%edi,%ebx,4) movhps %xmm4,24(%edi,%edx,4) ## Second hydrogen Z forces for 4 j particles ## Just load the four Z coords into one reg. each movss 32(%edi,%eax,4),%xmm4 movss 32(%edi,%ebx,4),%xmm5 movss 32(%edi,%ecx,4),%xmm6 movss 32(%edi,%edx,4),%xmm7 ## add what we have on the stack addss nb102_fjzH2(%esp),%xmm4 addss nb102_fjzH2b(%esp),%xmm5 addss nb102_fjzH2c(%esp),%xmm6 addss nb102_fjzH2d(%esp),%xmm7 ## store back movss %xmm4,32(%edi,%eax,4) movss %xmm5,32(%edi,%ebx,4) movss %xmm6,32(%edi,%ecx,4) movss %xmm7,32(%edi,%edx,4) ## should we do one more iteration? subl $4,nb102_innerk(%esp) jl _nb_kernel102_ia32_sse.nb102_single_check jmp _nb_kernel102_ia32_sse.nb102_unroll_loop_nb_kernel102_ia32_sse.nb102_single_check: addl $4,nb102_innerk(%esp) jnz _nb_kernel102_ia32_sse.nb102_single_loop jmp _nb_kernel102_ia32_sse.nb102_updateouterdata_nb_kernel102_ia32_sse.nb102_single_loop: movl nb102_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb102_innerjjnr(%esp) movl nb102_pos(%ebp),%esi leal (%eax,%eax,2),%eax ## fetch j coordinates xorps %xmm3,%xmm3 xorps %xmm4,%xmm4 xorps %xmm5,%xmm5 movss (%esi,%eax,4),%xmm3 ## jxO - - - movss 4(%esi,%eax,4),%xmm4 ## jyO - - - movss 8(%esi,%eax,4),%xmm5 ## jzO - - - 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 nb102_ixO(%esp),%xmm0 movaps nb102_iyO(%esp),%xmm1 movaps nb102_izO(%esp),%xmm2 movlhps %xmm6,%xmm3 ## xmm3 = jxO 0 jxH1 jxH2 shufps $228,%xmm6,%xmm4 ## constant 11100100 ;# xmm4 = jyO 0 jyH1 jyH2 shufps $68,%xmm7,%xmm5 ## constant 01000100 ;# xmm5 = jzO 0 jzH1 jzH2 ## store all j coordinates in jO movaps %xmm3,nb102_jxO(%esp) movaps %xmm4,nb102_jyO(%esp) movaps %xmm5,nb102_jzO(%esp) subps %xmm3,%xmm0 subps %xmm4,%xmm1 subps %xmm5,%xmm2 movaps %xmm0,nb102_dxOO(%esp) movaps %xmm1,nb102_dyOO(%esp) movaps %xmm2,nb102_dzOO(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 ## have rsq in xmm0 ## do invsqrt rsqrtps %xmm0,%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb102_three(%esp),%xmm3 mulps %xmm0,%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb102_half(%esp),%xmm3 ## rinv iO - j water xorps %xmm1,%xmm1 movaps %xmm3,%xmm0 xorps %xmm4,%xmm4 mulps %xmm0,%xmm0 ## xmm0=rinvsq ## fetch charges to xmm4 (temporary) movss nb102_qqOO(%esp),%xmm4 movhps nb102_qqOH(%esp),%xmm4 mulps %xmm4,%xmm3 ## xmm3=vcoul mulps %xmm3,%xmm0 ## total fscal addps nb102_vctot(%esp),%xmm3 movaps %xmm3,nb102_vctot(%esp) movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 mulps nb102_dxOO(%esp),%xmm0 mulps nb102_dyOO(%esp),%xmm1 mulps nb102_dzOO(%esp),%xmm2 ## initial update for j forces xorps %xmm3,%xmm3 xorps %xmm4,%xmm4 xorps %xmm5,%xmm5 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 movaps %xmm3,nb102_fjxO(%esp) movaps %xmm4,nb102_fjyO(%esp) movaps %xmm5,nb102_fjzO(%esp) addps nb102_fixO(%esp),%xmm0 addps nb102_fiyO(%esp),%xmm1 addps nb102_fizO(%esp),%xmm2 movaps %xmm0,nb102_fixO(%esp) movaps %xmm1,nb102_fiyO(%esp) movaps %xmm2,nb102_fizO(%esp) ## done with i O Now do i H1 & H2 simultaneously first get i particle coords: movaps nb102_ixH1(%esp),%xmm0 movaps nb102_iyH1(%esp),%xmm1 movaps nb102_izH1(%esp),%xmm2 movaps nb102_ixH2(%esp),%xmm3 movaps nb102_iyH2(%esp),%xmm4 movaps nb102_izH2(%esp),%xmm5 subps nb102_jxO(%esp),%xmm0 subps nb102_jyO(%esp),%xmm1 subps nb102_jzO(%esp),%xmm2 subps nb102_jxO(%esp),%xmm3 subps nb102_jyO(%esp),%xmm4 subps nb102_jzO(%esp),%xmm5 movaps %xmm0,nb102_dxH1O(%esp) movaps %xmm1,nb102_dyH1O(%esp) movaps %xmm2,nb102_dzH1O(%esp) movaps %xmm3,nb102_dxH2O(%esp) movaps %xmm4,nb102_dyH2O(%esp) movaps %xmm5,nb102_dzH2O(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm3,%xmm4 addps %xmm2,%xmm0 ## have rsqH1 in xmm0 addps %xmm5,%xmm4 ## have rsqH2 in xmm4 ## do invsqrt rsqrtps %xmm0,%xmm1 rsqrtps %xmm4,%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb102_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 nb102_half(%esp),%xmm3 ## rinv H1 - j water mulps nb102_half(%esp),%xmm7 ## rinv H2 - j water ## assemble charges in xmm6 xorps %xmm6,%xmm6 ## do coulomb interaction movaps %xmm3,%xmm0 movss nb102_qqOH(%esp),%xmm6 movaps %xmm7,%xmm4 movhps nb102_qqHH(%esp),%xmm6 mulps %xmm0,%xmm0 ## rinvsq mulps %xmm4,%xmm4 ## rinvsq mulps %xmm6,%xmm3 ## vcoul mulps %xmm6,%xmm7 ## vcoul movaps %xmm3,%xmm2 addps %xmm7,%xmm2 ## total vcoul mulps %xmm3,%xmm0 ## fscal addps nb102_vctot(%esp),%xmm2 mulps %xmm4,%xmm7 ## fscal movaps %xmm2,nb102_vctot(%esp) movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 mulps nb102_dxH1O(%esp),%xmm0 mulps nb102_dyH1O(%esp),%xmm1 mulps nb102_dzH1O(%esp),%xmm2⌨️ 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?