nb_kernel234_ia32_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 1,980 行 · 第 1/5 页

        movaps %xmm2,nb234_fizM(%esp)

        ## M-M interaction 
        movaps nb234_rinvMM(%esp),%xmm0
        movaps %xmm0,%xmm7      ## xmm7=rinv 
        movaps nb234_krf(%esp),%xmm5
        movaps %xmm0,%xmm1
        mulps  nb234_rsqMM(%esp),%xmm5   ## xmm5=krsq 
        movaps %xmm5,%xmm4
        addps  %xmm7,%xmm4      ## xmm4=r inv+ krsq 
        subps  nb234_crf(%esp),%xmm4
        mulps %xmm0,%xmm0
        mulps  nb234_qqMM(%esp),%xmm4   ## xmm4=voul=qq*(rinv+ krsq-crf) 
        mulps  nb234_two(%esp),%xmm5
        subps  %xmm5,%xmm7      ## xmm7=rinv-2*krsq 
        mulps  nb234_qqMM(%esp),%xmm7   ## xmm7 = coul part of fscal 
        addps  %xmm4,%xmm6      ## add to local vctot 
        mulps %xmm7,%xmm0       ## fsOH2 
        movaps %xmm0,%xmm1
        movaps %xmm0,%xmm2

        movaps %xmm0,%xmm1
        movaps %xmm6,nb234_vctot(%esp)
        movaps %xmm0,%xmm2

        movaps nb234_fjxM(%esp),%xmm3
        movaps nb234_fjyM(%esp),%xmm4
        movaps nb234_fjzM(%esp),%xmm5
        mulps nb234_dxMM(%esp),%xmm0
        mulps nb234_dyMM(%esp),%xmm1
        mulps nb234_dzMM(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb234_fixM(%esp),%xmm0
        addps nb234_fiyM(%esp),%xmm1
        addps nb234_fizM(%esp),%xmm2
        movaps %xmm3,nb234_fjxM(%esp)
        movaps %xmm4,nb234_fjyM(%esp)
        movaps %xmm5,nb234_fjzM(%esp)
        movaps %xmm0,nb234_fixM(%esp)
        movaps %xmm1,nb234_fiyM(%esp)
        movaps %xmm2,nb234_fizM(%esp)

        movl nb234_faction(%ebp),%edi
        ## update j forces 
        ## 4 j waters with four atoms each.
        ## step 1 : transpose fjxO, fjyO, fjzO, fjxH1
        movaps nb234_fjxO(%esp),%xmm0
        movaps nb234_fjyO(%esp),%xmm1
        movaps nb234_fjzO(%esp),%xmm2
        movaps nb234_fjxH1(%esp),%xmm3
        movaps %xmm0,%xmm4
        movaps %xmm1,%xmm5
        unpcklps %xmm2,%xmm4
        unpcklps %xmm3,%xmm5
        unpckhps %xmm2,%xmm0
        unpckhps %xmm3,%xmm1
        movaps %xmm4,%xmm2
        movaps %xmm0,%xmm3
        unpcklps %xmm5,%xmm4
        unpckhps %xmm5,%xmm2

        unpcklps %xmm1,%xmm0
        unpckhps %xmm1,%xmm3
        ## results are now in xmm4, xmm2, xmm0, xmm3
        ## load the corresponding j forces from memory
        movlps   (%edi,%eax,4),%xmm1
        movlps   (%edi,%ebx,4),%xmm5
        movlps   (%edi,%ecx,4),%xmm6
        movlps   (%edi,%edx,4),%xmm7
        movhps   8(%edi,%eax,4),%xmm1
        movhps   8(%edi,%ebx,4),%xmm5
        movhps   8(%edi,%ecx,4),%xmm6
        movhps   8(%edi,%edx,4),%xmm7
        ## add
        addps    %xmm4,%xmm1
        addps    %xmm2,%xmm5
        addps    %xmm0,%xmm6
        addps    %xmm3,%xmm7
        ## store back
        movlps   %xmm1,(%edi,%eax,4)
        movlps   %xmm5,(%edi,%ebx,4)
        movlps   %xmm6,(%edi,%ecx,4)
        movlps   %xmm7,(%edi,%edx,4)
        movhps   %xmm1,8(%edi,%eax,4)
        movhps   %xmm5,8(%edi,%ebx,4)
        movhps   %xmm6,8(%edi,%ecx,4)
        movhps   %xmm7,8(%edi,%edx,4)

        ## step 2 : transpose fjyH1, fjzH1, fjxH2, fjyH2
        movaps nb234_fjyH1(%esp),%xmm0
        movaps nb234_fjzH1(%esp),%xmm1
        movaps nb234_fjxH2(%esp),%xmm2
        movaps nb234_fjyH2(%esp),%xmm3
        movaps %xmm0,%xmm4
        movaps %xmm1,%xmm5
        unpcklps %xmm2,%xmm4
        unpcklps %xmm3,%xmm5
        unpckhps %xmm2,%xmm0
        unpckhps %xmm3,%xmm1
        movaps %xmm4,%xmm2
        movaps %xmm0,%xmm3
        unpcklps %xmm5,%xmm4
        unpckhps %xmm5,%xmm2

        unpcklps %xmm1,%xmm0
        unpckhps %xmm1,%xmm3
        ## results are now in xmm4, xmm2, xmm0, xmm3
        ## load the corresponding j forces from memory
        movlps   16(%edi,%eax,4),%xmm1
        movlps   16(%edi,%ebx,4),%xmm5
        movlps   16(%edi,%ecx,4),%xmm6
        movlps   16(%edi,%edx,4),%xmm7
        movhps   24(%edi,%eax,4),%xmm1
        movhps   24(%edi,%ebx,4),%xmm5
        movhps   24(%edi,%ecx,4),%xmm6
        movhps   24(%edi,%edx,4),%xmm7
        ## add
        addps    %xmm4,%xmm1
        addps    %xmm2,%xmm5
        addps    %xmm0,%xmm6
        addps    %xmm3,%xmm7
        ## store back
        movlps   %xmm1,16(%edi,%eax,4)
        movlps   %xmm5,16(%edi,%ebx,4)
        movlps   %xmm6,16(%edi,%ecx,4)
        movlps   %xmm7,16(%edi,%edx,4)
        movhps   %xmm1,24(%edi,%eax,4)
        movhps   %xmm5,24(%edi,%ebx,4)
        movhps   %xmm6,24(%edi,%ecx,4)
        movhps   %xmm7,24(%edi,%edx,4)

        ## step 3 : transpose fjzH2, fjxM, fjyM, fjzM. xmm4 is scratch
        movaps nb234_fjzH2(%esp),%xmm0
        movaps nb234_fjxM(%esp),%xmm1
        movaps nb234_fjyM(%esp),%xmm2
        movaps nb234_fjzM(%esp),%xmm3

        movaps %xmm0,%xmm4
        movaps %xmm1,%xmm5
        unpcklps %xmm2,%xmm4
        unpcklps %xmm3,%xmm5
        unpckhps %xmm2,%xmm0
        unpckhps %xmm3,%xmm1
        movaps %xmm4,%xmm2
        movaps %xmm0,%xmm3
        unpcklps %xmm5,%xmm4
        unpckhps %xmm5,%xmm2

        unpcklps %xmm1,%xmm0
        unpckhps %xmm1,%xmm3

        ## results are now in xmm0, xmm1, xmm2, xmm3
        ## load the corresponding j forces from memory
        movlps   32(%edi,%eax,4),%xmm1
        movlps   32(%edi,%ebx,4),%xmm5
        movlps   32(%edi,%ecx,4),%xmm6
        movlps   32(%edi,%edx,4),%xmm7
        movhps   40(%edi,%eax,4),%xmm1
        movhps   40(%edi,%ebx,4),%xmm5
        movhps   40(%edi,%ecx,4),%xmm6
        movhps   40(%edi,%edx,4),%xmm7
        ## add
        addps    %xmm4,%xmm1
        addps    %xmm2,%xmm5
        addps    %xmm0,%xmm6
        addps    %xmm3,%xmm7
        ## store back
        movlps   %xmm1,32(%edi,%eax,4)
        movlps   %xmm5,32(%edi,%ebx,4)
        movlps   %xmm6,32(%edi,%ecx,4)
        movlps   %xmm7,32(%edi,%edx,4)
        movhps   %xmm1,40(%edi,%eax,4)
        movhps   %xmm5,40(%edi,%ebx,4)
        movhps   %xmm6,40(%edi,%ecx,4)
        movhps   %xmm7,40(%edi,%edx,4)

        ## should we do one more iteration? 
        subl $4,nb234_innerk(%esp)
        jl    _nb_kernel234_ia32_sse.nb234_single_check
        jmp   _nb_kernel234_ia32_sse.nb234_unroll_loop
_nb_kernel234_ia32_sse.nb234_single_check: 
        addl $4,nb234_innerk(%esp)
        jnz   _nb_kernel234_ia32_sse.nb234_single_loop
        jmp   _nb_kernel234_ia32_sse.nb234_updateouterdata
_nb_kernel234_ia32_sse.nb234_single_loop: 
        movl  nb234_innerjjnr(%esp),%edx        ## pointer to jjnr[k] 
        movl  (%edx),%eax
        addl $4,nb234_innerjjnr(%esp)

        movl nb234_pos(%ebp),%esi
        leal  (%eax,%eax,2),%eax

        ## fetch j coordinates
        movlps (%esi,%eax,4),%xmm3              ##  Ox  Oy  
        movlps 16(%esi,%eax,4),%xmm4            ## H1y H1z 
        movlps 32(%esi,%eax,4),%xmm5            ## H2z  Mx 
        movhps 8(%esi,%eax,4),%xmm3             ##  Ox  Oy  Oz H1x
        movhps 24(%esi,%eax,4),%xmm4            ## H1y H1z H2x H2y
        movhps 40(%esi,%eax,4),%xmm5            ## H2z  Mx  My  Mz
        ## transpose
        movaps %xmm4,%xmm0
        movaps %xmm3,%xmm1
        movaps %xmm4,%xmm2
        movaps %xmm3,%xmm6
        shufps $18,%xmm5,%xmm4 ## (00010010)  h2x - Mx  - 
        shufps $193,%xmm0,%xmm3 ## (11000001)  Oy  - H1y - 
        shufps $35,%xmm5,%xmm2 ## (00100011) H2y - My  - 
        shufps $18,%xmm0,%xmm1 ## (00010010)  Oz  - H1z - 
        ##  xmm6: Ox - - H1x   xmm5: H2z - - Mz 
        shufps $140,%xmm4,%xmm6 ## (10001100) Ox H1x H2x Mx 
        shufps $136,%xmm2,%xmm3 ## (10001000) Oy H1y H2y My 
        shufps $200,%xmm5,%xmm1 ## (11001000) Oz H1z H2z Mz

        ## store all j coordinates in jO  
        movaps %xmm6,nb234_jxO(%esp)
        movaps %xmm3,nb234_jyO(%esp)
        movaps %xmm1,nb234_jzO(%esp)

        ## do O and M in parallel
        movaps nb234_ixO(%esp),%xmm0
        movaps nb234_iyO(%esp),%xmm1
        movaps nb234_izO(%esp),%xmm2
        movaps nb234_ixM(%esp),%xmm3
        movaps nb234_iyM(%esp),%xmm4
        movaps nb234_izM(%esp),%xmm5
        subps  nb234_jxO(%esp),%xmm0
        subps  nb234_jyO(%esp),%xmm1
        subps  nb234_jzO(%esp),%xmm2
        subps  nb234_jxO(%esp),%xmm3
        subps  nb234_jyO(%esp),%xmm4
        subps  nb234_jzO(%esp),%xmm5

        movaps %xmm0,nb234_dxOO(%esp)
        movaps %xmm1,nb234_dyOO(%esp)
        movaps %xmm2,nb234_dzOO(%esp)
        movaps %xmm3,nb234_dxMM(%esp)
        movaps %xmm4,nb234_dyMM(%esp)
        movaps %xmm5,nb234_dzMM(%esp)

        mulps %xmm0,%xmm0
        mulps %xmm1,%xmm1
        mulps %xmm2,%xmm2
        addps %xmm1,%xmm0
        addps %xmm2,%xmm0       ## have rsq in xmm0 
        mulps %xmm3,%xmm3
        mulps %xmm4,%xmm4
        mulps %xmm5,%xmm5
        addps %xmm3,%xmm4
        addps %xmm5,%xmm4       ## have rsq in xmm4
        ## Save data 
        movaps %xmm0,nb234_rsqOO(%esp)
        movaps %xmm4,nb234_rsqMM(%esp)

        ## do 1/x for O
        rsqrtps %xmm0,%xmm1
        movaps  %xmm1,%xmm2
        mulps   %xmm1,%xmm1
        movaps  nb234_three(%esp),%xmm3
        mulps   %xmm0,%xmm1     ## rsq*lu*lu 
        subps   %xmm1,%xmm3     ## constant 30-rsq*lu*lu 
        mulps   %xmm2,%xmm3     ## lu*(3-rsq*lu*lu) 
        mulps   nb234_half(%esp),%xmm3
        movaps  %xmm3,nb234_rinvOO(%esp)        ## rinvH2 

        ## 1/sqrt(x) for M
        rsqrtps %xmm4,%xmm5
        movaps  %xmm5,%xmm6
        mulps   %xmm5,%xmm5
        movaps  nb234_three(%esp),%xmm7
        mulps   %xmm4,%xmm5
        subps   %xmm5,%xmm7
        mulps   %xmm6,%xmm7
        mulps   nb234_half(%esp),%xmm7   ## rinv iH1 - j water 
        movaps %xmm7,nb234_rinvMM(%esp)


        ## LJ table interaction
        movaps nb234_rinvOO(%esp),%xmm0
        movss %xmm0,%xmm1
        mulss  nb234_rsqOO(%esp),%xmm1   ## xmm1=r 
        mulss  nb234_tsc(%esp),%xmm1

    cvttps2pi %xmm1,%mm6
    cvtpi2ps %mm6,%xmm3
        subss    %xmm3,%xmm1    ## xmm1=eps 
    movss %xmm1,%xmm2
    mulss  %xmm2,%xmm2      ## xmm2=eps2 
    pslld $3,%mm6

    movd %eax,%mm0

    movl nb234_VFtab(%ebp),%esi
    movd %mm6,%eax

    ## dispersion 
    movlps (%esi,%eax,4),%xmm5
    movaps %xmm5,%xmm4
    shufps $136,%xmm7,%xmm4 ## constant 10001000
    shufps $221,%xmm7,%xmm5 ## constant 11011101

    movlps 8(%esi,%eax,4),%xmm7
    movaps %xmm7,%xmm6
    shufps $136,%xmm3,%xmm6 ## constant 10001000
    shufps $221,%xmm3,%xmm7 ## constant 11011101
    ## dispersion table ready, in xmm4-xmm7 
    mulss  %xmm1,%xmm6      ## xmm6=Geps 
    mulss  %xmm2,%xmm7      ## xmm7=Heps2 
    addss  %xmm6,%xmm5
    addss  %xmm7,%xmm5      ## xmm5=Fp 
    mulss  nb234_two(%esp),%xmm7         ## two*Heps2 
    addss  %xmm6,%xmm7
    addss  %xmm5,%xmm7 ## xmm7=FF 
    mulss  %xmm1,%xmm5 ## xmm5=eps*Fp 
    addss  %xmm4,%xmm5 ## xmm5=VV 

    movss nb234_c6(%esp),%xmm4
    mulss  %xmm4,%xmm7   ## fijD 
    mulss  %xmm4,%xmm5   ## Vvdw6 
        xorps  %xmm3,%xmm3

        mulps  nb234_tsc(%esp),%xmm7
        subss  %xmm7,%xmm3
        movss  %xmm3,nb234_fstmp(%esp)

    addss  nb234_Vvdwtot(%esp),%xmm5
    movss %xmm5,nb234_Vvdwtot(%esp)

    ## repulsion 
    movlps 16(%esi,%eax,4),%xmm5
    movaps %xmm5,%xmm4
    shufps $136,%xmm7,%xmm4 ## constant 10001000
    shufps $221,%xmm7,%xmm5 ## constant 11011101

    movlps 24(%esi,%eax,4),%xmm7
    movaps %xmm7,%xmm6
    shufps $136,%xmm3,%xmm6 ## constant 10001000
    shufps $221,%xmm3,%xmm7 ## constant 11011101
    ## table ready, in xmm4-xmm7 
    mulss  %xmm1,%xmm6      ## xmm6=Geps 
    mulss  %xmm2,%xmm7      ## xmm7=Heps2 
    addss  %xmm6,%xmm5
    addss  %xmm7,%xmm5      ## xmm5=Fp 
    mulss  nb234_two(%esp),%xmm7         ## two*Heps2 
    addss  %xmm6,%xmm7
    addss  %xmm5,%xmm7 ## xmm7=FF 
    mulss  %xmm1,%xmm5 ## xmm5=eps*Fp 
    addss  %xmm4,%xmm5 ## xmm5=VV 

    movss nb234_c12(%esp),%xmm4
    mulss  %xmm4,%xmm7 ## fijR 
    mulss  %xmm4,%xmm5 ## Vvdw12 
        movaps nb234_fstmp(%esp),%xmm3
        mulss  nb234_tsc(%esp),%xmm7
        subss  %xmm7,%xmm3

    addss  nb234_Vvdwtot(%esp),%xmm5
    movss %xmm5,nb234_Vvdwtot(%esp)

        mulss %xmm3,%xmm0
        movaps %xmm0,%xmm1
        movaps %xmm0,%xmm2

        movd  %mm0,%eax

        mulss  nb234_dxOO(%esp),%xmm0
        mulss  nb234_dyOO(%esp),%xmm1
        mulss  nb234_dzOO(%esp),%xmm2
        xorps   %xmm3,%xmm3
        xorps   %xmm4,%xmm4
        xorps   %xmm5,%xmm5
        subss   %xmm0,%xmm3
        subss   %xmm1,%xmm4
        subss   %xmm2,%xmm5
        movaps  %xmm3,nb234_fjxO(%esp)
        movaps  %xmm4,nb234_fjyO(%esp)
        movaps  %xmm5,nb234_fjzO(%esp)
        addss   nb234_fixO(%esp),%xmm0
        addss   nb234_fiyO(%esp),%xmm1
        addss   nb234_fizO(%esp),%xmm2
        movss  %xmm0,nb234_fixO(%esp)
        movss  %xmm1,nb234_fiyO(%esp)
        movss  %xmm2,nb234_fizO(%esp)

        ## do  M coulomb interaction
        movaps nb234_rinvMM(%esp),%xmm0
        movaps %xmm0,%xmm7      ## xmm7=rinv 
        movaps nb234_krf(%esp),%xmm5
        mulps  %xmm0,%xmm0      ## xmm0=rinvsq 

        ## fetch charges to xmm3 (temporary) 
        xorps  %xmm3,%xmm3
        movss   nb234_qqMH(%esp),%xmm3
        movhps  nb234_qqMM(%esp),%xmm3
        shufps $193,%xmm3,%xmm3 ## constant 11000001