nb_kernel114_ia32_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,017 行 · 第 1/5 页

        movaps nb114_ixO(%esp),%xmm0
        movaps nb114_iyO(%esp),%xmm1
        movaps nb114_izO(%esp),%xmm2
        movaps nb114_ixM(%esp),%xmm3
        movaps nb114_iyM(%esp),%xmm4
        movaps nb114_izM(%esp),%xmm5
        subps  nb114_jxO(%esp),%xmm0
        subps  nb114_jyO(%esp),%xmm1
        subps  nb114_jzO(%esp),%xmm2
        subps  nb114_jxO(%esp),%xmm3
        subps  nb114_jyO(%esp),%xmm4
        subps  nb114_jzO(%esp),%xmm5

        movaps %xmm0,nb114_dxOO(%esp)
        movaps %xmm1,nb114_dyOO(%esp)
        movaps %xmm2,nb114_dzOO(%esp)
        movaps %xmm3,nb114_dxMM(%esp)
        movaps %xmm4,nb114_dyMM(%esp)
        movaps %xmm5,nb114_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 M data 
        movaps %xmm4,nb114_rsqMM(%esp)

        ## do 1/x for O and 1/sqrt(x) for M
        rcpss  %xmm0,%xmm1
        rsqrtps %xmm4,%xmm5
        movss  nb114_two(%esp),%xmm2
        movaps  %xmm5,%xmm6
        mulss  %xmm1,%xmm0
        mulps   %xmm5,%xmm5
        subss  %xmm0,%xmm2
        movaps  nb114_three(%esp),%xmm7
        mulss  %xmm1,%xmm2      ## constant 1/r2

        mulps   %xmm4,%xmm5
        movss  %xmm2,%xmm0
        subps   %xmm5,%xmm7
        mulss  %xmm2,%xmm2
        mulps   %xmm6,%xmm7
        mulss  %xmm0,%xmm2      ## constant 1/r6
        mulps   nb114_half(%esp),%xmm7   ## rinv iH1 - j water 
        movss  %xmm2,%xmm1
        movaps %xmm7,nb114_rinvMM(%esp)

        mulss  %xmm2,%xmm2      ## constant 1/r12
        mulss  nb114_c6(%esp),%xmm1
        mulss  nb114_c12(%esp),%xmm2
        movss  %xmm2,%xmm3
        subss  %xmm1,%xmm3
        addss  nb114_Vvdwtot(%esp),%xmm3
        movss  %xmm3,nb114_Vvdwtot(%esp)
        mulss  nb114_six(%esp),%xmm1
        mulss  nb114_twelve(%esp),%xmm2
        subss  %xmm1,%xmm2
        mulss  %xmm2,%xmm0      ## fscal
        movss  %xmm0,%xmm1
        movss  %xmm0,%xmm2
        mulss  nb114_dxOO(%esp),%xmm0
        mulss  nb114_dyOO(%esp),%xmm1
        mulss  nb114_dzOO(%esp),%xmm2
        xorps   %xmm3,%xmm3
        xorps   %xmm4,%xmm4
        xorps   %xmm5,%xmm5
        subss   %xmm0,%xmm3
        subss   %xmm1,%xmm4
        subss   %xmm2,%xmm5
        movaps  %xmm3,nb114_fjxO(%esp)
        movaps  %xmm4,nb114_fjyO(%esp)
        movaps  %xmm5,nb114_fjzO(%esp)
        addss   nb114_fixO(%esp),%xmm0
        addss   nb114_fiyO(%esp),%xmm1
        addss   nb114_fizO(%esp),%xmm2
        movss  %xmm0,nb114_fixO(%esp)
        movss  %xmm1,nb114_fiyO(%esp)
        movss  %xmm2,nb114_fizO(%esp)

        ## do  M coulomb interaction
        movaps nb114_rinvMM(%esp),%xmm0
        movaps %xmm0,%xmm4      ## xmm4=rinv
        mulps  %xmm0,%xmm0      ## xmm0=rinvsq 

        ## fetch charges to xmm3 (temporary) 
        xorps  %xmm3,%xmm3
        movss   nb114_qqMH(%esp),%xmm3
        movhps  nb114_qqMM(%esp),%xmm3
        shufps $193,%xmm3,%xmm3 ## constant 11000001 

        mulps  %xmm3,%xmm4      ## xmm4=voul    
        mulps  %xmm4,%xmm0
        addps  nb114_vctot(%esp),%xmm4
        movaps %xmm4,nb114_vctot(%esp)

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

        mulps   nb114_dxMM(%esp),%xmm0
        mulps   nb114_dyMM(%esp),%xmm1
        mulps   nb114_dzMM(%esp),%xmm2
        ## update forces M - j water 
        movaps  nb114_fjxO(%esp),%xmm3
        movaps  nb114_fjyO(%esp),%xmm4
        movaps  nb114_fjzO(%esp),%xmm5
        subps   %xmm0,%xmm3
        subps   %xmm1,%xmm4
        subps   %xmm2,%xmm5
        movaps  %xmm3,nb114_fjxO(%esp)
        movaps  %xmm4,nb114_fjyO(%esp)
        movaps  %xmm5,nb114_fjzO(%esp)
        addps   nb114_fixM(%esp),%xmm0
        addps   nb114_fiyM(%esp),%xmm1
        addps   nb114_fizM(%esp),%xmm2
        movaps  %xmm0,nb114_fixM(%esp)
        movaps  %xmm1,nb114_fiyM(%esp)
        movaps  %xmm2,nb114_fizM(%esp)

        ## i H1 & H2 simultaneously first get i particle coords: 
        movaps  nb114_ixH1(%esp),%xmm0
        movaps  nb114_iyH1(%esp),%xmm1
        movaps  nb114_izH1(%esp),%xmm2
        movaps  nb114_ixH2(%esp),%xmm3
        movaps  nb114_iyH2(%esp),%xmm4
        movaps  nb114_izH2(%esp),%xmm5
        subps   nb114_jxO(%esp),%xmm0
        subps   nb114_jyO(%esp),%xmm1
        subps   nb114_jzO(%esp),%xmm2
        subps   nb114_jxO(%esp),%xmm3
        subps   nb114_jyO(%esp),%xmm4
        subps   nb114_jzO(%esp),%xmm5
        movaps %xmm0,nb114_dxH1H1(%esp)
        movaps %xmm1,nb114_dyH1H1(%esp)
        movaps %xmm2,nb114_dzH1H1(%esp)
        movaps %xmm3,nb114_dxH2H2(%esp)
        movaps %xmm4,nb114_dyH2H2(%esp)
        movaps %xmm5,nb114_dzH2H2(%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 
        movaps  %xmm0,nb114_rsqH1H1(%esp)
        movaps  %xmm4,nb114_rsqH2H2(%esp)

        ## start doing invsqrt use rsq values in xmm0, xmm4 
        rsqrtps %xmm0,%xmm1
        rsqrtps %xmm4,%xmm5
        movaps  %xmm1,%xmm2
        movaps  %xmm5,%xmm6
        mulps   %xmm1,%xmm1
        mulps   %xmm5,%xmm5
        movaps  nb114_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   nb114_half(%esp),%xmm3   ## rinvH1H1
        mulps   nb114_half(%esp),%xmm7   ## rinvH2H2
        movaps  %xmm3,nb114_rinvH1H1(%esp)
        movaps  %xmm7,nb114_rinvH2H2(%esp)

        ## Do H1 coulomb interaction
        movaps nb114_rinvH1H1(%esp),%xmm0
        movaps %xmm0,%xmm4      ## xmm4=rinv 
        mulps  %xmm0,%xmm0      ## xmm0=rinvsq 

        ## fetch charges to xmm3 (temporary) 
        xorps  %xmm3,%xmm3
        movss   nb114_qqHH(%esp),%xmm3
        movhps  nb114_qqMH(%esp),%xmm3
        shufps $193,%xmm3,%xmm3 ## constant 11000001 

        mulps  %xmm3,%xmm4      ## xmm4=voul
        mulps  %xmm4,%xmm0
        addps  nb114_vctot(%esp),%xmm4
        movaps %xmm4,nb114_vctot(%esp)


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

        mulps   nb114_dxH1H1(%esp),%xmm0
        mulps   nb114_dyH1H1(%esp),%xmm1
        mulps   nb114_dzH1H1(%esp),%xmm2
        ## update forces H1 - j water 
        movaps  nb114_fjxO(%esp),%xmm3
        movaps  nb114_fjyO(%esp),%xmm4
        movaps  nb114_fjzO(%esp),%xmm5
        subps   %xmm0,%xmm3
        subps   %xmm1,%xmm4
        subps   %xmm2,%xmm5
        movaps  %xmm3,nb114_fjxO(%esp)
        movaps  %xmm4,nb114_fjyO(%esp)
        movaps  %xmm5,nb114_fjzO(%esp)
        addps   nb114_fixH1(%esp),%xmm0
        addps   nb114_fiyH1(%esp),%xmm1
        addps   nb114_fizH1(%esp),%xmm2
        movaps  %xmm0,nb114_fixH1(%esp)
        movaps  %xmm1,nb114_fiyH1(%esp)
        movaps  %xmm2,nb114_fizH1(%esp)

        ## H2 Coulomb
        movaps nb114_rinvH2H2(%esp),%xmm0
        movaps %xmm0,%xmm4      ## xmm4=rinv 
        mulps  %xmm0,%xmm0      ## xmm0=rinvsq 

        ## fetch charges to xmm3 (temporary) 
        xorps  %xmm3,%xmm3
        movss   nb114_qqHH(%esp),%xmm3
        movhps  nb114_qqMH(%esp),%xmm3
        shufps $193,%xmm3,%xmm3 ## constant 11000001 

        mulps  %xmm3,%xmm4      ## xmm4=voul
        mulps  %xmm4,%xmm0
        addps  nb114_vctot(%esp),%xmm4   ## local vctot summation variable
        movaps %xmm4,nb114_vctot(%esp)

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

        mulps   nb114_dxH2H2(%esp),%xmm0
        mulps   nb114_dyH2H2(%esp),%xmm1
        mulps   nb114_dzH2H2(%esp),%xmm2
        ## update forces H2 - j water 
        movaps  nb114_fjxO(%esp),%xmm3
        movaps  nb114_fjyO(%esp),%xmm4
        movaps  nb114_fjzO(%esp),%xmm5
        subps   %xmm0,%xmm3
        subps   %xmm1,%xmm4
        subps   %xmm2,%xmm5
        movaps  %xmm3,nb114_fjxO(%esp)
        movaps  %xmm4,nb114_fjyO(%esp)
        movaps  %xmm5,nb114_fjzO(%esp)
        addps   nb114_fixH2(%esp),%xmm0
        addps   nb114_fiyH2(%esp),%xmm1
        addps   nb114_fizH2(%esp),%xmm2
        movaps  %xmm0,nb114_fixH2(%esp)
        movaps  %xmm1,nb114_fiyH2(%esp)
        movaps  %xmm2,nb114_fizH2(%esp)

        movl    nb114_faction(%ebp),%esi
        ## update j water forces from local variables.
        ## transpose back first
        movaps  nb114_fjxO(%esp),%xmm0   ## Ox H1x H2x Mx 
        movaps  nb114_fjyO(%esp),%xmm1   ## Oy H1y H2y My
        movaps  nb114_fjzO(%esp),%xmm2   ## Oz H1z H2z Mz

        movaps  %xmm0,%xmm3
        movaps  %xmm0,%xmm4
        unpcklps %xmm1,%xmm3            ## Ox Oy - -
        shufps $0x1,%xmm2,%xmm4        ## h1x - Oz -
        movaps  %xmm1,%xmm5
        movaps  %xmm0,%xmm6
        unpcklps %xmm2,%xmm5            ## - - H1y H1z
        unpckhps %xmm1,%xmm6            ## h2x h2y - - 
        unpckhps %xmm2,%xmm1            ## - - My Mz

        shufps  $0x32,%xmm0,%xmm2 ## (00110010) h2z - Mx -
        shufps  $36,%xmm4,%xmm3 ## constant 00100100 ;# Ox Oy Oz H1x 
        shufps  $78,%xmm6,%xmm5 ## constant 01001110 ;# h1y h1z h2x h2y
        shufps  $232,%xmm1,%xmm2 ## constant 11101000 ;# h2z mx my mz

        movlps  (%esi,%eax,4),%xmm0
        movlps  16(%esi,%eax,4),%xmm1
        movlps  32(%esi,%eax,4),%xmm4
        movhps  8(%esi,%eax,4),%xmm0
        movhps  24(%esi,%eax,4),%xmm1
        movhps  40(%esi,%eax,4),%xmm4
        addps   %xmm3,%xmm0
        addps   %xmm5,%xmm1
        addps   %xmm2,%xmm4
        movlps   %xmm0,(%esi,%eax,4)
        movlps   %xmm1,16(%esi,%eax,4)
        movlps   %xmm4,32(%esi,%eax,4)
        movhps   %xmm0,8(%esi,%eax,4)
        movhps   %xmm1,24(%esi,%eax,4)
        movhps   %xmm4,40(%esi,%eax,4)

        decl nb114_innerk(%esp)
        jz    _nb_kernel114_ia32_sse.nb114_updateouterdata
        jmp   _nb_kernel114_ia32_sse.nb114_single_loop
_nb_kernel114_ia32_sse.nb114_updateouterdata: 
        movl  nb114_ii3(%esp),%ecx
        movl  nb114_faction(%ebp),%edi
        movl  nb114_fshift(%ebp),%esi
        movl  nb114_is3(%esp),%edx

        ## accumulate  Oi forces in xmm0, xmm1, xmm2 
        movaps nb114_fixO(%esp),%xmm0
        movaps nb114_fiyO(%esp),%xmm1
        movaps nb114_fizO(%esp),%xmm2

        movhlps %xmm0,%xmm3
        movhlps %xmm1,%xmm4
        movhlps %xmm2,%xmm5
        addps  %xmm3,%xmm0
        addps  %xmm4,%xmm1
        addps  %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 

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

        shufps $1,%xmm3,%xmm3
        shufps $1,%xmm4,%xmm4
        shufps $1,%xmm5,%xmm5
        addss  %xmm3,%xmm0
        addss  %xmm4,%xmm1
        addss  %xmm5,%xmm2      ## xmm0-xmm2 has single force in pos0 

        ## increment i force 
        movss  (%edi,%ecx,4),%xmm3
        movss  4(%edi,%ecx,4),%xmm4
        movss  8(%edi,%ecx,4),%xmm5
        addss  %xmm0,%xmm3
        addss  %xmm1,%xmm4
        addss  %xmm2,%xmm5
        movss  %xmm3,(%edi,%ecx,4)
        movss  %xmm4,4(%edi,%ecx,4)
        movss  %xmm5,8(%edi,%ecx,4)

        ## accumulate force in xmm6/xmm7 for fshift 
        movaps %xmm0,%xmm6
        movss %xmm2,%xmm7
        movlhps %xmm1,%xmm6
        shufps $8,%xmm6,%xmm6 ## constant 00001000      

        ## accumulate H1i forces in xmm0, xmm1, xmm2 
        movaps nb114_fixH1(%esp),%xmm0
        movaps nb114_fiyH1(%esp),%xmm1
        movaps nb114_fizH1(%esp),%xmm2

        movhlps %xmm0,%xmm3
        movhlps %xmm1,%xmm4
        movhlps %xmm2,%xmm5
        addps  %xmm3,%xmm0
        addps  %xmm4,%xmm1
        addps  %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 

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

        shufps $1,%xmm3,%xmm3
        shufps $1,%xmm4,%xmm4
        shufps $1,%xmm5,%xmm5
        addss  %xmm3,%xmm0
        addss  %xmm4,%xmm1
        addss  %xmm5,%xmm2      ## xmm0-xmm2 has single force in pos0 

        ## increment i force 
        movss  12(%edi,%ecx,4),%xmm3
        movss  16(%edi,%ecx,4),%xmm4
        movss  20(%edi,%ecx,4),%xmm5
        addss  %xmm0,%xmm3
        addss  %xmm1,%xmm4
        addss  %xmm2,%xmm5
        movss  %xmm3,12(%edi,%ecx,4)
        movss  %xmm4,16(%edi,%ecx,4)
        movss  %xmm5,20(%edi,%ecx,4)

        ## accumulate force in xmm6/xmm7 for fshift 
        addss %xmm2,%xmm7
        movlhps %xmm1,%xmm0
        shufps $8,%xmm0,%xmm0 ## constant 00001000      
        addps   %xmm0,%xmm6

        ## accumulate H2i forces in xmm0, xmm1, xmm2 
        movaps nb114_fixH2(%esp),%xmm0
        movaps nb114_fiyH2(%esp),%xmm1
        movaps nb114_fizH2(%esp),%xmm2

        movhlps %xmm0,%xmm3
        movhlps %xmm1,%xmm4
        movhlps %xmm2,%xmm5
        addps  %xmm3,%xmm0
        addps  %xmm4,%xmm1
        addps  %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 

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

        shufps $1,%xmm3,%xmm3
        shufps $1,%xmm4,%xmm4
        shufps $1,%xmm5,%xmm5
        addss  %xmm3,%xmm0
        addss  %xmm4,%xmm1