nb_kernel302_ia32_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 1,970 行 · 第 1/5 页

        movaps %xmm3,nb302_fjxH2(%esp)
        movaps %xmm4,nb302_fjyH2(%esp)
        movaps %xmm5,nb302_fjzH2(%esp)
        movaps %xmm0,nb302_fixH1(%esp)
        movaps %xmm1,nb302_fiyH1(%esp)
        movaps %xmm2,nb302_fizH1(%esp)

        ## H2-O interaction 
        movaps nb302_rinvH2O(%esp),%xmm0
        movaps %xmm0,%xmm1
        mulps  nb302_rsqH2O(%esp),%xmm1   ## xmm1=r 
        mulps  nb302_tsc(%esp),%xmm1
        movhlps %xmm1,%xmm2
        cvttps2pi %xmm1,%mm6
        cvttps2pi %xmm2,%mm7    ## mm6/mm7 contain lu indices 
        cvtpi2ps %mm6,%xmm3
        cvtpi2ps %mm7,%xmm2
        movlhps  %xmm2,%xmm3
        subps    %xmm3,%xmm1    ## xmm1=eps 
        movaps %xmm1,%xmm2
        mulps  %xmm2,%xmm2      ## xmm2=eps2 

        pslld   $2,%mm6
        pslld   $2,%mm7

        movd %mm6,%eax
        psrlq $32,%mm6
        movd %mm7,%ecx
        psrlq $32,%mm7
        movd %mm6,%ebx
        movd %mm7,%edx

        movlps (%esi,%eax,4),%xmm5
        movlps (%esi,%ecx,4),%xmm7
        movhps (%esi,%ebx,4),%xmm5
        movhps (%esi,%edx,4),%xmm7 ## got half coulomb table 

        movaps %xmm5,%xmm4
        shufps $136,%xmm7,%xmm4 ## constant 10001000
        shufps $221,%xmm7,%xmm5 ## constant 11011101

        movlps 8(%esi,%eax,4),%xmm7
        movlps 8(%esi,%ecx,4),%xmm3
        movhps 8(%esi,%ebx,4),%xmm7
        movhps 8(%esi,%edx,4),%xmm3    ## other half of coulomb table  
        movaps %xmm7,%xmm6
        shufps $136,%xmm3,%xmm6 ## constant 10001000
        shufps $221,%xmm3,%xmm7 ## constant 11011101
        ## coulomb table ready, in xmm4-xmm7  

        mulps  %xmm1,%xmm6      ## xmm6=Geps 
        mulps  %xmm2,%xmm7      ## xmm7=Heps2 
        addps  %xmm6,%xmm5
        addps  %xmm7,%xmm5      ## xmm5=Fp 
        mulps  nb302_two(%esp),%xmm7            ## two*Heps2 
        movaps nb302_qqOH(%esp),%xmm3
        addps  %xmm6,%xmm7
        addps  %xmm5,%xmm7 ## xmm7=FF 
        mulps  %xmm1,%xmm5 ## xmm5=eps*Fp 
        addps  %xmm4,%xmm5 ## xmm5=VV 
        mulps  %xmm3,%xmm5 ## vcoul=qq*VV  
        mulps  %xmm7,%xmm3 ## fijC=FF*qq 
        ## at this point mm5 contains vcoul and mm3 fijC 

        addps  nb302_vctot(%esp),%xmm5
        movaps %xmm5,nb302_vctot(%esp)
        xorps  %xmm1,%xmm1
        mulps  nb302_tsc(%esp),%xmm3
        mulps  %xmm0,%xmm3
        subps  %xmm3,%xmm1

        movaps %xmm1,%xmm0
        movaps %xmm1,%xmm2

        movaps nb302_fjxO(%esp),%xmm3
        movaps nb302_fjyO(%esp),%xmm4
        movaps nb302_fjzO(%esp),%xmm5
        mulps nb302_dxH2O(%esp),%xmm0
        mulps nb302_dyH2O(%esp),%xmm1
        mulps nb302_dzH2O(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb302_fixH2(%esp),%xmm0
        addps nb302_fiyH2(%esp),%xmm1
        addps nb302_fizH2(%esp),%xmm2
        movaps %xmm3,nb302_fjxO(%esp)
        movaps %xmm4,nb302_fjyO(%esp)
        movaps %xmm5,nb302_fjzO(%esp)
        movaps %xmm0,nb302_fixH2(%esp)
        movaps %xmm1,nb302_fiyH2(%esp)
        movaps %xmm2,nb302_fizH2(%esp)

        ## H2-H1 interaction 
        movaps nb302_rinvH2H1(%esp),%xmm0
        movaps %xmm0,%xmm1
        mulps  nb302_rsqH2H1(%esp),%xmm1   ## xmm1=r 
        mulps  nb302_tsc(%esp),%xmm1
        movhlps %xmm1,%xmm2
        cvttps2pi %xmm1,%mm6
        cvttps2pi %xmm2,%mm7    ## mm6/mm7 contain lu indices 
        cvtpi2ps %mm6,%xmm3
        cvtpi2ps %mm7,%xmm2
        movlhps  %xmm2,%xmm3
        subps    %xmm3,%xmm1    ## xmm1=eps 
        movaps %xmm1,%xmm2
        mulps  %xmm2,%xmm2      ## xmm2=eps2 

        pslld   $2,%mm6
        pslld   $2,%mm7

        movd %mm6,%eax
        psrlq $32,%mm6
        movd %mm7,%ecx
        psrlq $32,%mm7
        movd %mm6,%ebx
        movd %mm7,%edx

        movlps (%esi,%eax,4),%xmm5
        movlps (%esi,%ecx,4),%xmm7
        movhps (%esi,%ebx,4),%xmm5
        movhps (%esi,%edx,4),%xmm7 ## got half coulomb table 

        movaps %xmm5,%xmm4
        shufps $136,%xmm7,%xmm4 ## constant 10001000
        shufps $221,%xmm7,%xmm5 ## constant 11011101

        movlps 8(%esi,%eax,4),%xmm7
        movlps 8(%esi,%ecx,4),%xmm3
        movhps 8(%esi,%ebx,4),%xmm7
        movhps 8(%esi,%edx,4),%xmm3    ## other half of coulomb table  
        movaps %xmm7,%xmm6
        shufps $136,%xmm3,%xmm6 ## constant 10001000
        shufps $221,%xmm3,%xmm7 ## constant 11011101
        ## coulomb table ready, in xmm4-xmm7  

        mulps  %xmm1,%xmm6      ## xmm6=Geps 
        mulps  %xmm2,%xmm7      ## xmm7=Heps2 
        addps  %xmm6,%xmm5
        addps  %xmm7,%xmm5      ## xmm5=Fp 
        mulps  nb302_two(%esp),%xmm7            ## two*Heps2 
        movaps nb302_qqHH(%esp),%xmm3
        addps  %xmm6,%xmm7
        addps  %xmm5,%xmm7 ## xmm7=FF 
        mulps  %xmm1,%xmm5 ## xmm5=eps*Fp 
        addps  %xmm4,%xmm5 ## xmm5=VV 
        mulps  %xmm3,%xmm5 ## vcoul=qq*VV  
        mulps  %xmm7,%xmm3 ## fijC=FF*qq 
        ## at this point mm5 contains vcoul and mm3 fijC 

        addps  nb302_vctot(%esp),%xmm5
        movaps %xmm5,nb302_vctot(%esp)
        xorps  %xmm1,%xmm1
        mulps  nb302_tsc(%esp),%xmm3
        mulps  %xmm0,%xmm3
        subps  %xmm3,%xmm1

        movaps %xmm1,%xmm0
        movaps %xmm1,%xmm2

        movaps nb302_fjxH1(%esp),%xmm3
        movaps nb302_fjyH1(%esp),%xmm4
        movaps nb302_fjzH1(%esp),%xmm5
        mulps nb302_dxH2H1(%esp),%xmm0
        mulps nb302_dyH2H1(%esp),%xmm1
        mulps nb302_dzH2H1(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb302_fixH2(%esp),%xmm0
        addps nb302_fiyH2(%esp),%xmm1
        addps nb302_fizH2(%esp),%xmm2
        movaps %xmm3,nb302_fjxH1(%esp)
        movaps %xmm4,nb302_fjyH1(%esp)
        movaps %xmm5,nb302_fjzH1(%esp)
        movaps %xmm0,nb302_fixH2(%esp)
        movaps %xmm1,nb302_fiyH2(%esp)
        movaps %xmm2,nb302_fizH2(%esp)

        ## H2-H2 interaction 
        movaps nb302_rinvH2H2(%esp),%xmm0
        movaps %xmm0,%xmm1
        mulps  nb302_rsqH2H2(%esp),%xmm1   ## xmm1=r 
        mulps  nb302_tsc(%esp),%xmm1
        movhlps %xmm1,%xmm2
        cvttps2pi %xmm1,%mm6
        cvttps2pi %xmm2,%mm7    ## mm6/mm7 contain lu indices 
        cvtpi2ps %mm6,%xmm3
        cvtpi2ps %mm7,%xmm2
        movlhps  %xmm2,%xmm3
        subps    %xmm3,%xmm1    ## xmm1=eps 
        movaps %xmm1,%xmm2
        mulps  %xmm2,%xmm2      ## xmm2=eps2 

        pslld   $2,%mm6
        pslld   $2,%mm7

        movd %mm6,%eax
        psrlq $32,%mm6
        movd %mm7,%ecx
        psrlq $32,%mm7
        movd %mm6,%ebx
        movd %mm7,%edx

        movlps (%esi,%eax,4),%xmm5
        movlps (%esi,%ecx,4),%xmm7
        movhps (%esi,%ebx,4),%xmm5
        movhps (%esi,%edx,4),%xmm7 ## got half coulomb table 

        movaps %xmm5,%xmm4
        shufps $136,%xmm7,%xmm4 ## constant 10001000
        shufps $221,%xmm7,%xmm5 ## constant 11011101

        movlps 8(%esi,%eax,4),%xmm7
        movlps 8(%esi,%ecx,4),%xmm3
        movhps 8(%esi,%ebx,4),%xmm7
        movhps 8(%esi,%edx,4),%xmm3    ## other half of coulomb table  
        movaps %xmm7,%xmm6
        shufps $136,%xmm3,%xmm6 ## constant 10001000
        shufps $221,%xmm3,%xmm7 ## constant 11011101
        ## coulomb table ready, in xmm4-xmm7  

        mulps  %xmm1,%xmm6      ## xmm6=Geps 
        mulps  %xmm2,%xmm7      ## xmm7=Heps2 
        addps  %xmm6,%xmm5
        addps  %xmm7,%xmm5      ## xmm5=Fp 
        mulps  nb302_two(%esp),%xmm7            ## two*Heps2 
        movaps nb302_qqHH(%esp),%xmm3
        addps  %xmm6,%xmm7
        addps  %xmm5,%xmm7 ## xmm7=FF 
        mulps  %xmm1,%xmm5 ## xmm5=eps*Fp 
        addps  %xmm4,%xmm5 ## xmm5=VV 
        mulps  %xmm3,%xmm5 ## vcoul=qq*VV  
        mulps  %xmm7,%xmm3 ## fijC=FF*qq 
        ## at this point mm5 contains vcoul and mm3 fijC 

        addps  nb302_vctot(%esp),%xmm5
        movaps %xmm5,nb302_vctot(%esp)
        xorps  %xmm1,%xmm1
        mulps  nb302_tsc(%esp),%xmm3
        mulps  %xmm0,%xmm3
        subps  %xmm3,%xmm1

        movaps %xmm1,%xmm0
        movaps %xmm1,%xmm2

        movaps nb302_fjxH2(%esp),%xmm3
        movaps nb302_fjyH2(%esp),%xmm4
        movaps nb302_fjzH2(%esp),%xmm5
        mulps nb302_dxH2H2(%esp),%xmm0
        mulps nb302_dyH2H2(%esp),%xmm1
        mulps nb302_dzH2H2(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb302_fixH2(%esp),%xmm0
        addps nb302_fiyH2(%esp),%xmm1
        addps nb302_fizH2(%esp),%xmm2
        movaps %xmm3,nb302_fjxH2(%esp)
        movaps %xmm4,nb302_fjyH2(%esp)
        movaps %xmm5,nb302_fjzH2(%esp)
        movaps %xmm0,nb302_fixH2(%esp)
        movaps %xmm1,nb302_fiyH2(%esp)
        movaps %xmm2,nb302_fizH2(%esp)

        movl nb302_faction(%ebp),%edi

        movd %mm0,%eax
        movd %mm1,%ebx
        movd %mm2,%ecx
        movd %mm3,%edx

        ## 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 nb302_fjxO(%esp),%xmm0   ## xmm0= fjxOa  fjxOb  fjxOc  fjxOd 
        movaps nb302_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 nb302_fjzO(%esp),%xmm0    ## xmm0= fjzOa   fjzOb   fjzOc   fjzOd 
        movaps nb302_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 nb302_fjyH1(%esp),%xmm0    ## xmm0= fjyH1a  fjyH1b  fjyH1c  fjyH1d 
        movaps nb302_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 nb302_fjxH2(%esp),%xmm0    ## xmm0= fjxH2a  fjxH2b  fjxH2c  fjxH2d 
        movaps nb302_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 nb302_fjzH2(%esp),%xmm4
        addss nb302_fjzH2b(%esp),%xmm5
        addss nb302_fjzH2c(%esp),%xmm6
        addss nb302_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,nb302_innerk(%esp)
        jl    _nb_kernel302_ia32_sse.nb302_single_check
        jmp   _nb_kernel302_ia32_sse.nb302_unroll_loop
_nb_kernel302_ia32_sse.nb302_single_check: 
        addl $4,nb302_innerk(%esp)
        jnz   _nb_kernel302_ia32_sse.nb302_single_loop
        jmp   _nb_kernel302_ia32_sse.nb302_updateouterdata
_nb_kernel302_ia32_sse.nb302_single_loop: 
        movl  nb302_innerjjnr(%esp),%edx        ## pointer to jjnr[k] 
        movl  (%edx),%eax
        addl $4,nb302_innerjjnr(%esp)

        movl nb302_pos(%ebp),%esi
        leal  (%eax,%eax,2),%eax

        ## fetch j coordinates 
        xorps %xmm3,%xmm3
        xorps %xmm4,%xmm4