nb_kernel132_ia32_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,018 行 · 第 1/5 页

        xorps %xmm3,%xmm3
        movaps %xmm3,%xmm4
        movaps %xmm3,%xmm5
        mulps nb132_dxOO(%esp),%xmm0
        mulps nb132_dyOO(%esp),%xmm1
        mulps nb132_dzOO(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb132_fixO(%esp),%xmm0
        addps nb132_fiyO(%esp),%xmm1
        addps nb132_fizO(%esp),%xmm2
        movaps %xmm3,nb132_fjxO(%esp)
        movaps %xmm4,nb132_fjyO(%esp)
        movaps %xmm5,nb132_fjzO(%esp)
        movaps %xmm0,nb132_fixO(%esp)
        movaps %xmm1,nb132_fiyO(%esp)
        movaps %xmm2,nb132_fizO(%esp)

        ## O-H1 interaction 
        movaps nb132_rinvOH1(%esp),%xmm0
        movaps %xmm0,%xmm4      ## xmm7=rinv 
        mulps  %xmm0,%xmm0  ## rinvsq
        mulps  nb132_qqOH(%esp),%xmm4   ## xmm4=voul=qq*(rinv+ krsq) 
        mulps  %xmm4,%xmm0

        addps  %xmm4,%xmm6      ## add to local vctot 
        movaps %xmm0,%xmm1
        movaps %xmm0,%xmm2

        xorps %xmm3,%xmm3
        movaps %xmm3,%xmm4
        movaps %xmm3,%xmm5
        mulps nb132_dxOH1(%esp),%xmm0
        mulps nb132_dyOH1(%esp),%xmm1
        mulps nb132_dzOH1(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb132_fixO(%esp),%xmm0
        addps nb132_fiyO(%esp),%xmm1
        addps nb132_fizO(%esp),%xmm2
        movaps %xmm3,nb132_fjxH1(%esp)
        movaps %xmm4,nb132_fjyH1(%esp)
        movaps %xmm5,nb132_fjzH1(%esp)
        movaps %xmm0,nb132_fixO(%esp)
        movaps %xmm1,nb132_fiyO(%esp)
        movaps %xmm2,nb132_fizO(%esp)

        ## O-H2 interaction  
        movaps nb132_rinvOH2(%esp),%xmm0
        movaps %xmm0,%xmm4      ## xmm7=rinv 
        mulps  %xmm0,%xmm0  ## rinvsq
        mulps  nb132_qqOH(%esp),%xmm4   ## xmm4=voul=qq*(rinv+ krsq) 
        mulps  %xmm4,%xmm0

        addps  %xmm4,%xmm6      ## add to local vctot 
        movaps %xmm0,%xmm1
        movaps %xmm0,%xmm2

        xorps %xmm3,%xmm3
        movaps %xmm3,%xmm4
        movaps %xmm3,%xmm5
        mulps nb132_dxOH2(%esp),%xmm0
        mulps nb132_dyOH2(%esp),%xmm1
        mulps nb132_dzOH2(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb132_fixO(%esp),%xmm0
        addps nb132_fiyO(%esp),%xmm1
        addps nb132_fizO(%esp),%xmm2
        movaps %xmm3,nb132_fjxH2(%esp)
        movaps %xmm4,nb132_fjyH2(%esp)
        movaps %xmm5,nb132_fjzH2(%esp)
        movaps %xmm0,nb132_fixO(%esp)
        movaps %xmm1,nb132_fiyO(%esp)
        movaps %xmm2,nb132_fizO(%esp)

        ## H1-O interaction 
        movaps nb132_rinvH1O(%esp),%xmm0
        movaps %xmm0,%xmm4      ## xmm7=rinv 
        mulps  %xmm0,%xmm0  ## rinvsq
        mulps  nb132_qqOH(%esp),%xmm4   ## xmm4=voul=qq*(rinv+ krsq) 
        mulps  %xmm4,%xmm0

        addps  %xmm4,%xmm6      ## add to local vctot 
        movaps %xmm0,%xmm1
        movaps %xmm0,%xmm2

        movaps nb132_fjxO(%esp),%xmm3
        movaps nb132_fjyO(%esp),%xmm4
        movaps nb132_fjzO(%esp),%xmm5
        mulps nb132_dxH1O(%esp),%xmm0
        mulps nb132_dyH1O(%esp),%xmm1
        mulps nb132_dzH1O(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb132_fixH1(%esp),%xmm0
        addps nb132_fiyH1(%esp),%xmm1
        addps nb132_fizH1(%esp),%xmm2
        movaps %xmm3,nb132_fjxO(%esp)
        movaps %xmm4,nb132_fjyO(%esp)
        movaps %xmm5,nb132_fjzO(%esp)
        movaps %xmm0,nb132_fixH1(%esp)
        movaps %xmm1,nb132_fiyH1(%esp)
        movaps %xmm2,nb132_fizH1(%esp)

        ## H1-H1 interaction 
        movaps nb132_rinvH1H1(%esp),%xmm0
        movaps %xmm0,%xmm4      ## xmm7=rinv 
        mulps  %xmm0,%xmm0  ## rinvsq
        mulps  nb132_qqHH(%esp),%xmm4   ## xmm4=voul=qq*(rinv+ krsq) 
        mulps  %xmm4,%xmm0

        addps  %xmm4,%xmm6      ## add to local vctot 
        movaps %xmm0,%xmm1
        movaps %xmm0,%xmm2

        movaps nb132_fjxH1(%esp),%xmm3
        movaps nb132_fjyH1(%esp),%xmm4
        movaps nb132_fjzH1(%esp),%xmm5
        mulps nb132_dxH1H1(%esp),%xmm0
        mulps nb132_dyH1H1(%esp),%xmm1
        mulps nb132_dzH1H1(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb132_fixH1(%esp),%xmm0
        addps nb132_fiyH1(%esp),%xmm1
        addps nb132_fizH1(%esp),%xmm2
        movaps %xmm3,nb132_fjxH1(%esp)
        movaps %xmm4,nb132_fjyH1(%esp)
        movaps %xmm5,nb132_fjzH1(%esp)
        movaps %xmm0,nb132_fixH1(%esp)
        movaps %xmm1,nb132_fiyH1(%esp)
        movaps %xmm2,nb132_fizH1(%esp)

        ## H1-H2 interaction 
        movaps nb132_rinvH1H2(%esp),%xmm0
        movaps %xmm0,%xmm4      ## xmm7=rinv 
        mulps  %xmm0,%xmm0  ## rinvsq
        mulps  nb132_qqHH(%esp),%xmm4   ## xmm4=voul=qq*(rinv+ krsq) 
        mulps  %xmm4,%xmm0

        addps  %xmm4,%xmm6      ## add to local vctot 
        movaps %xmm0,%xmm1
        movaps %xmm0,%xmm2

        movaps nb132_fjxH2(%esp),%xmm3
        movaps nb132_fjyH2(%esp),%xmm4
        movaps nb132_fjzH2(%esp),%xmm5
        mulps nb132_dxH1H2(%esp),%xmm0
        mulps nb132_dyH1H2(%esp),%xmm1
        mulps nb132_dzH1H2(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb132_fixH1(%esp),%xmm0
        addps nb132_fiyH1(%esp),%xmm1
        addps nb132_fizH1(%esp),%xmm2
        movaps %xmm3,nb132_fjxH2(%esp)
        movaps %xmm4,nb132_fjyH2(%esp)
        movaps %xmm5,nb132_fjzH2(%esp)
        movaps %xmm0,nb132_fixH1(%esp)
        movaps %xmm1,nb132_fiyH1(%esp)
        movaps %xmm2,nb132_fizH1(%esp)

        ## H2-O interaction 
        movaps nb132_rinvH2O(%esp),%xmm0
        movaps %xmm0,%xmm4      ## xmm7=rinv 
        mulps  %xmm0,%xmm0  ## rinvsq
        mulps  nb132_qqOH(%esp),%xmm4   ## xmm4=voul=qq*(rinv+ krsq) 
        mulps  %xmm4,%xmm0

        addps  %xmm4,%xmm6      ## add to local vctot 
        movaps %xmm0,%xmm1
        movaps %xmm0,%xmm2

        movaps nb132_fjxO(%esp),%xmm3
        movaps nb132_fjyO(%esp),%xmm4
        movaps nb132_fjzO(%esp),%xmm5
        mulps nb132_dxH2O(%esp),%xmm0
        mulps nb132_dyH2O(%esp),%xmm1
        mulps nb132_dzH2O(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb132_fixH2(%esp),%xmm0
        addps nb132_fiyH2(%esp),%xmm1
        addps nb132_fizH2(%esp),%xmm2
        movaps %xmm3,nb132_fjxO(%esp)
        movaps %xmm4,nb132_fjyO(%esp)
        movaps %xmm5,nb132_fjzO(%esp)
        movaps %xmm0,nb132_fixH2(%esp)
        movaps %xmm1,nb132_fiyH2(%esp)
        movaps %xmm2,nb132_fizH2(%esp)

        ## H2-H1 interaction 
        movaps nb132_rinvH2H1(%esp),%xmm0
        movaps %xmm0,%xmm4      ## xmm7=rinv 
        mulps  %xmm0,%xmm0  ## rinvsq
        mulps  nb132_qqHH(%esp),%xmm4   ## xmm4=voul=qq*(rinv+ krsq) 
        mulps  %xmm4,%xmm0

        addps  %xmm4,%xmm6      ## add to local vctot 
        movaps %xmm0,%xmm1
        movaps %xmm0,%xmm2

        movaps nb132_fjxH1(%esp),%xmm3
        movaps nb132_fjyH1(%esp),%xmm4
        movaps nb132_fjzH1(%esp),%xmm5
        mulps nb132_dxH2H1(%esp),%xmm0
        mulps nb132_dyH2H1(%esp),%xmm1
        mulps nb132_dzH2H1(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb132_fixH2(%esp),%xmm0
        addps nb132_fiyH2(%esp),%xmm1
        addps nb132_fizH2(%esp),%xmm2
        movaps %xmm3,nb132_fjxH1(%esp)
        movaps %xmm4,nb132_fjyH1(%esp)
        movaps %xmm5,nb132_fjzH1(%esp)
        movaps %xmm0,nb132_fixH2(%esp)
        movaps %xmm1,nb132_fiyH2(%esp)
        movaps %xmm2,nb132_fizH2(%esp)

        ## H2-H2 interaction 
        movaps nb132_rinvH2H2(%esp),%xmm0
        movaps %xmm0,%xmm4      ## xmm7=rinv 
        mulps  %xmm0,%xmm0  ## rinvsq
        mulps  nb132_qqHH(%esp),%xmm4   ## xmm4=voul=qq*(rinv+ krsq) 
        mulps  %xmm4,%xmm0

        addps  %xmm4,%xmm6      ## add to local vctot 
        movaps %xmm0,%xmm1
        movaps %xmm0,%xmm2

        movaps %xmm0,%xmm1
        movaps %xmm6,nb132_vctot(%esp)
        movaps %xmm0,%xmm2

        movaps nb132_fjxH2(%esp),%xmm3
        movaps nb132_fjyH2(%esp),%xmm4
        movaps nb132_fjzH2(%esp),%xmm5
        mulps nb132_dxH2H2(%esp),%xmm0
        mulps nb132_dyH2H2(%esp),%xmm1
        mulps nb132_dzH2H2(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb132_fixH2(%esp),%xmm0
        addps nb132_fiyH2(%esp),%xmm1
        addps nb132_fizH2(%esp),%xmm2
        movaps %xmm3,nb132_fjxH2(%esp)
        movaps %xmm4,nb132_fjyH2(%esp)
        movaps %xmm5,nb132_fjzH2(%esp)
        movaps %xmm0,nb132_fixH2(%esp)
        movaps %xmm1,nb132_fiyH2(%esp)
        movaps %xmm2,nb132_fizH2(%esp)

        movl nb132_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 nb132_fjxO(%esp),%xmm0   ## xmm0= fjxOa  fjxOb  fjxOc  fjxOd 
        movaps nb132_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 nb132_fjzO(%esp),%xmm0    ## xmm0= fjzOa   fjzOb   fjzOc   fjzOd 
        movaps nb132_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 nb132_fjyH1(%esp),%xmm0    ## xmm0= fjyH1a  fjyH1b  fjyH1c  fjyH1d 
        movaps nb132_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 nb132_fjxH2(%esp),%xmm0    ## xmm0= fjxH2a  fjxH2b  fjxH2c  fjxH2d 
        movaps nb132_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 nb132_fjzH2(%esp),%xmm4
        addss nb132_fjzH2b(%esp),%xmm5
        addss nb132_fjzH2c(%esp),%xmm6
        addss nb132_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,nb132_innerk(%esp)
        jl    _nb_kernel132_ia32_sse.nb132_single_check
        jmp   _nb_kernel132_ia32_sse.nb132_unroll_loop
_nb_kernel132_ia32_sse.nb132_single_check: 
        addl $4,nb132_innerk(%esp)
        jnz   _nb_kernel132_ia32_sse.nb132_single_loop
        jmp   _nb_kernel132_ia32_sse.nb132_updateouterdata
_nb_kernel132_ia32_sse.nb132_single_loop: 
        movl  nb132_innerjjnr(%esp),%edx       ## pointer to jjnr[k] 
        movl  (%edx),%eax
        addl $4,nb132_innerjjnr(%esp)

        movl nb132_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  nb132_ixO(%esp),%xmm0