nb_kernel104_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,968 行 · 第 1/5 页
S
1,968 行
subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb104_fixH2(%esp),%xmm0 addps nb104_fiyH2(%esp),%xmm1 addps nb104_fizH2(%esp),%xmm2 movaps %xmm3,nb104_fjxH2(%esp) movaps %xmm4,nb104_fjyH2(%esp) movaps %xmm5,nb104_fjzH2(%esp) movaps %xmm0,nb104_fixH2(%esp) movaps %xmm1,nb104_fiyH2(%esp) movaps %xmm2,nb104_fizH2(%esp) ## H2-M interaction movaps nb104_rinvH2M(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb104_qqMH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs H2-M addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb104_fjxM(%esp),%xmm3 movaps nb104_fjyM(%esp),%xmm4 movaps nb104_fjzM(%esp),%xmm5 mulps nb104_dxH2M(%esp),%xmm0 mulps nb104_dyH2M(%esp),%xmm1 mulps nb104_dzH2M(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb104_fixH2(%esp),%xmm0 addps nb104_fiyH2(%esp),%xmm1 addps nb104_fizH2(%esp),%xmm2 movaps %xmm3,nb104_fjxM(%esp) movaps %xmm4,nb104_fjyM(%esp) movaps %xmm5,nb104_fjzM(%esp) movaps %xmm0,nb104_fixH2(%esp) movaps %xmm1,nb104_fiyH2(%esp) movaps %xmm2,nb104_fizH2(%esp) ## M-H1 interaction movaps nb104_rinvMH1(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb104_qqMH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs M-H1 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb104_fjxH1(%esp),%xmm3 movaps nb104_fjyH1(%esp),%xmm4 movaps nb104_fjzH1(%esp),%xmm5 mulps nb104_dxMH1(%esp),%xmm0 mulps nb104_dyMH1(%esp),%xmm1 mulps nb104_dzMH1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb104_fixM(%esp),%xmm0 addps nb104_fiyM(%esp),%xmm1 addps nb104_fizM(%esp),%xmm2 movaps %xmm3,nb104_fjxH1(%esp) movaps %xmm4,nb104_fjyH1(%esp) movaps %xmm5,nb104_fjzH1(%esp) movaps %xmm0,nb104_fixM(%esp) movaps %xmm1,nb104_fiyM(%esp) movaps %xmm2,nb104_fizM(%esp) ## M-H2 interaction movaps nb104_rinvMH2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb104_qqMH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs M-H2 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb104_fjxH2(%esp),%xmm3 movaps nb104_fjyH2(%esp),%xmm4 movaps nb104_fjzH2(%esp),%xmm5 mulps nb104_dxMH2(%esp),%xmm0 mulps nb104_dyMH2(%esp),%xmm1 mulps nb104_dzMH2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb104_fixM(%esp),%xmm0 addps nb104_fiyM(%esp),%xmm1 addps nb104_fizM(%esp),%xmm2 movaps %xmm3,nb104_fjxH2(%esp) movaps %xmm4,nb104_fjyH2(%esp) movaps %xmm5,nb104_fjzH2(%esp) movaps %xmm0,nb104_fixM(%esp) movaps %xmm1,nb104_fiyM(%esp) movaps %xmm2,nb104_fizM(%esp) ## M-M interaction movaps nb104_rinvMM(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb104_qqMM(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs M-M addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm7,nb104_vctot(%esp) movaps %xmm0,%xmm2 movaps nb104_fjxM(%esp),%xmm3 movaps nb104_fjyM(%esp),%xmm4 movaps nb104_fjzM(%esp),%xmm5 mulps nb104_dxMM(%esp),%xmm0 mulps nb104_dyMM(%esp),%xmm1 mulps nb104_dzMM(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb104_fixM(%esp),%xmm0 addps nb104_fiyM(%esp),%xmm1 addps nb104_fizM(%esp),%xmm2 movaps %xmm3,nb104_fjxM(%esp) movaps %xmm4,nb104_fjyM(%esp) movaps %xmm5,nb104_fjzM(%esp) movaps %xmm0,nb104_fixM(%esp) movaps %xmm1,nb104_fiyM(%esp) movaps %xmm2,nb104_fizM(%esp) movl nb104_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 nb104_fjxH1(%esp),%xmm0 ## xmm0= fjxH1a fjxH1b fjxH1c fjxH1d movaps nb104_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 nb104_fjzH1(%esp),%xmm0 ## xmm0= fjzH1a fjzH1b fjzH1c fjzH1d movaps nb104_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 nb104_fjyH2(%esp),%xmm0 ## xmm0= fjyH2a fjyH2b fjyH2c fjyH2d movaps nb104_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 nb104_fjxM(%esp),%xmm0 ## xmm0= fjxMa fjxMb fjxMc fjxMd movaps nb104_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 nb104_fjzM(%esp),%xmm4 addss nb104_fjzMb(%esp),%xmm5 addss nb104_fjzMc(%esp),%xmm6 addss nb104_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,nb104_innerk(%esp) jl _nb_kernel104_ia32_sse.nb104_single_check jmp _nb_kernel104_ia32_sse.nb104_unroll_loop_nb_kernel104_ia32_sse.nb104_single_check: addl $4,nb104_innerk(%esp) jnz _nb_kernel104_ia32_sse.nb104_single_loop jmp _nb_kernel104_ia32_sse.nb104_updateouterdata_nb_kernel104_ia32_sse.nb104_single_loop: movl nb104_innerjjnr(%esp),%edx ## pointer to jjnr[k] movl (%edx),%eax addl $4,nb104_innerjjnr(%esp) movl nb104_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 nb104_ixM(%esp),%xmm0 movaps nb104_iyM(%esp),%xmm1 movaps nb104_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,nb104_jxM(%esp) movaps %xmm4,nb104_jyM(%esp) movaps %xmm5,nb104_jzM(%esp) subps %xmm3,%xmm0 subps %xmm4,%xmm1 subps %xmm5,%xmm2 movaps %xmm0,nb104_dxMM(%esp) movaps %xmm1,nb104_dyMM(%esp) movaps %xmm2,nb104_dzMM(%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 nb104_three(%esp),%xmm3 mulps %xmm0,%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb104_half(%esp),%xmm3 ## rinv iM- j water xorps %xmm1,%xmm1 movaps %xmm3,%xmm0 xorps %xmm4,%xmm4 mulps %xmm0,%xmm0 ## xmm0=rinvsq ## fetch charges to xmm4 movss nb104_qqMM(%esp),%xmm4 movhps nb104_qqMH(%esp),%xmm4 mulps %xmm4,%xmm3 ## xmm3=vcoul mulps %xmm3,%xmm0 ## total fscal addps nb104_vctot(%esp),%xmm3 movaps %xmm3,nb104_vctot(%esp) movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 mulps nb104_dxMM(%esp),%xmm0 mulps nb104_dyMM(%esp),%xmm1 mulps nb104_dzMM(%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,nb104_fjxM(%esp) movaps %xmm4,nb104_fjyM(%esp) movaps %xmm5,nb104_fjzM(%esp) addps nb104_fixM(%esp),%xmm0 addps nb104_fiyM(%esp),%xmm1 addps nb104_fizM(%esp),%xmm2 movaps %xmm0,nb104_fixM(%esp) movaps %xmm1,nb104_fiyM(%esp) movaps %xmm2,nb104_fizM(%esp) ## done with i M Now do i H1 & H2 simultaneously first get i particle coords: movaps nb104_ixH1(%esp),%xmm0 movaps nb104_iyH1(%esp),%xmm1 movaps nb104_izH1(%esp),%xmm2 movaps nb104_ixH2(%esp),%xmm3 movaps nb104_iyH2(%esp),%xmm4 movaps nb104_izH2(%esp),%xmm5 subps nb104_jxM(%esp),%xmm0 subps nb104_jyM(%esp),%xmm1 subps nb104_jzM(%esp),%xmm2 subps nb104_jxM(%esp),%xmm3 subps nb104_jyM(%esp),%xmm4 subps nb104_jzM(%esp),%xmm5 movaps %xmm0,nb104_dxH1M(%esp) movaps %xmm1,nb104_dyH1M(%esp) movaps %xmm2,nb104_dzH1M(%esp) movaps %xmm3,nb104_dxH2M(%esp) movaps %xmm4,nb104_dyH2M(%esp) movaps %xmm5,nb104_dzH2M(%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 nb104_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 nb104_half(%esp),%xmm3 ## rinv H1 - j water mulps nb104_half(%esp),%xmm7 ## rinv H2 - j water ## assemble charges in xmm6 xorps %xmm6,%xmm6 movss nb104_qqMH(%esp),%xmm6 movhps nb104_qqHH(%esp),%xmm6 ## do coulomb interaction movaps %xmm3,%xmm0 movaps %xmm7,%xmm4 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 ⌨️ 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?