nb_kernel304_ia32_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 1,976 行 · 第 1/5 页

        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  nb304_two(%esp),%xmm7            ## two*Heps2 
        movaps nb304_qqMH(%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  nb304_vctot(%esp),%xmm5
        movaps %xmm5,nb304_vctot(%esp)
        xorps  %xmm1,%xmm1
        mulps  nb304_tsc(%esp),%xmm3
        mulps  %xmm0,%xmm3
        subps  %xmm3,%xmm1

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

        movaps nb304_fjxH1(%esp),%xmm3
        movaps nb304_fjyH1(%esp),%xmm4
        movaps nb304_fjzH1(%esp),%xmm5
        mulps nb304_dxMH1(%esp),%xmm0
        mulps nb304_dyMH1(%esp),%xmm1
        mulps nb304_dzMH1(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb304_fixM(%esp),%xmm0
        addps nb304_fiyM(%esp),%xmm1
        addps nb304_fizM(%esp),%xmm2
        movaps %xmm3,nb304_fjxH1(%esp)
        movaps %xmm4,nb304_fjyH1(%esp)
        movaps %xmm5,nb304_fjzH1(%esp)
        movaps %xmm0,nb304_fixM(%esp)
        movaps %xmm1,nb304_fiyM(%esp)
        movaps %xmm2,nb304_fizM(%esp)

        ## M-H2 interaction 
        movaps nb304_rinvMH2(%esp),%xmm0
        movaps %xmm0,%xmm1
        mulps  nb304_rsqMH2(%esp),%xmm1   ## xmm1=r 
        mulps  nb304_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  nb304_two(%esp),%xmm7            ## two*Heps2 
        movaps nb304_qqMH(%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  nb304_vctot(%esp),%xmm5
        movaps %xmm5,nb304_vctot(%esp)
        xorps  %xmm1,%xmm1
        mulps  nb304_tsc(%esp),%xmm3
        mulps  %xmm0,%xmm3
        subps  %xmm3,%xmm1

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

        movaps nb304_fjxH2(%esp),%xmm3
        movaps nb304_fjyH2(%esp),%xmm4
        movaps nb304_fjzH2(%esp),%xmm5
        mulps nb304_dxMH2(%esp),%xmm0
        mulps nb304_dyMH2(%esp),%xmm1
        mulps nb304_dzMH2(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb304_fixM(%esp),%xmm0
        addps nb304_fiyM(%esp),%xmm1
        addps nb304_fizM(%esp),%xmm2
        movaps %xmm3,nb304_fjxH2(%esp)
        movaps %xmm4,nb304_fjyH2(%esp)
        movaps %xmm5,nb304_fjzH2(%esp)
        movaps %xmm0,nb304_fixM(%esp)
        movaps %xmm1,nb304_fiyM(%esp)
        movaps %xmm2,nb304_fizM(%esp)

        ## M-M interaction 
        movaps nb304_rinvMM(%esp),%xmm0
        movaps %xmm0,%xmm1
        mulps  nb304_rsqMM(%esp),%xmm1   ## xmm1=r 
        mulps  nb304_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  nb304_two(%esp),%xmm7            ## two*Heps2 
        movaps nb304_qqMM(%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  nb304_vctot(%esp),%xmm5
        movaps %xmm5,nb304_vctot(%esp)
        xorps  %xmm1,%xmm1
        mulps  nb304_tsc(%esp),%xmm3
        mulps  %xmm0,%xmm3
        subps  %xmm3,%xmm1

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

        movaps nb304_fjxM(%esp),%xmm3
        movaps nb304_fjyM(%esp),%xmm4
        movaps nb304_fjzM(%esp),%xmm5
        mulps nb304_dxMM(%esp),%xmm0
        mulps nb304_dyMM(%esp),%xmm1
        mulps nb304_dzMM(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb304_fixM(%esp),%xmm0
        addps nb304_fiyM(%esp),%xmm1
        addps nb304_fizM(%esp),%xmm2
        movaps %xmm3,nb304_fjxM(%esp)
        movaps %xmm4,nb304_fjyM(%esp)
        movaps %xmm5,nb304_fjzM(%esp)
        movaps %xmm0,nb304_fixM(%esp)
        movaps %xmm1,nb304_fiyM(%esp)
        movaps %xmm2,nb304_fizM(%esp)

        movl nb304_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:
        ## 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 nb304_fjxH1(%esp),%xmm0   ## xmm0= fjxH1a  fjxH1b  fjxH1c  fjxH1d 
        movaps nb304_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 nb304_fjzH1(%esp),%xmm0    ## xmm0= fjzH1a   fjzH1b   fjzH1c   fjzH1d 
        movaps nb304_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 nb304_fjyH2(%esp),%xmm0   ## xmm0= fjyH2a  fjyH2b  fjyH2c  fjyH2d 
        movaps nb304_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 nb304_fjxM(%esp),%xmm0   ## xmm0= fjxMa  fjxMb  fjxMc  fjxMd 
        movaps nb304_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 nb304_fjzM(%esp),%xmm4
        addss nb304_fjzMb(%esp),%xmm5
        addss nb304_fjzMc(%esp),%xmm6
        addss nb304_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,nb304_innerk(%esp)
        jl    _nb_kernel304_ia32_sse.nb304_single_check
        jmp   _nb_kernel304_ia32_sse.nb304_unroll_loop
_nb_kernel304_ia32_sse.nb304_single_check: 
        addl $4,nb304_innerk(%esp)
        jnz   _nb_kernel304_ia32_sse.nb304_single_loop
        jmp   _nb_kernel304_ia32_sse.nb304_updateouterdata
_nb_kernel304_ia32_sse.nb304_single_loop: 
        movl  nb304_innerjjnr(%esp),%edx        ## pointer to jjnr[k] 
        movl  (%edx),%eax
        addl $4,nb304_innerjjnr(%esp)

        movl nb304_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  nb304_ixM(%esp),%xmm0
        movaps  nb304_iyM(%esp),%xmm1
        movaps  nb304_izM(%esp),%xmm2
        movlhps %xmm6,%xmm3                     ## xmm3 = jxM   0   jxH1 jxH2