nb_kernel134_x86_64_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,165 行 · 第 1/5 页

        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,nb134_jxO(%rsp)
        movaps %xmm3,nb134_jyO(%rsp)
        movaps %xmm1,nb134_jzO(%rsp)

    movaps %xmm6,%xmm0  ## jxO
    movaps %xmm1,%xmm2  ## jzO
    movaps %xmm3,%xmm1  ## jyO
    movaps %xmm3,%xmm4  ## jyO
    movaps %xmm6,%xmm3  ## jxO
    movaps %xmm2,%xmm5  ## jzO

        ## do O and M in parallel
        subps  nb134_ixO(%rsp),%xmm0
        subps  nb134_iyO(%rsp),%xmm1
        subps  nb134_izO(%rsp),%xmm2
        subps  nb134_ixM(%rsp),%xmm3
        subps  nb134_iyM(%rsp),%xmm4
        subps  nb134_izM(%rsp),%xmm5

        movaps %xmm0,nb134_dxOO(%rsp)
        movaps %xmm1,nb134_dyOO(%rsp)
        movaps %xmm2,nb134_dzOO(%rsp)
        movaps %xmm3,nb134_dxMM(%rsp)
        movaps %xmm4,nb134_dyMM(%rsp)
        movaps %xmm5,nb134_dzMM(%rsp)

        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,nb134_rsqOO(%rsp)
        movaps %xmm4,nb134_rsqMM(%rsp)

        ## do 1/x for O
        rsqrtps %xmm0,%xmm1
        movaps  %xmm1,%xmm2
        mulps   %xmm1,%xmm1
        movaps  nb134_three(%rsp),%xmm3
        mulps   %xmm0,%xmm1     ## rsq*lu*lu 
        subps   %xmm1,%xmm3     ## constant 30-rsq*lu*lu 
        mulps   %xmm2,%xmm3     ## lu*(3-rsq*lu*lu) 
        mulps   nb134_half(%rsp),%xmm3
        movaps  %xmm3,nb134_rinvOO(%rsp)        ## rinvH2 

        ## 1/sqrt(x) for M
        rsqrtps %xmm4,%xmm5
        movaps  %xmm5,%xmm6
        mulps   %xmm5,%xmm5
        movaps  nb134_three(%rsp),%xmm7
        mulps   %xmm4,%xmm5
        subps   %xmm5,%xmm7
        mulps   %xmm6,%xmm7
        mulps   nb134_half(%rsp),%xmm7   ## rinv iH1 - j water 
        movaps %xmm7,nb134_rinvMM(%rsp)


        ## LJ table interaction
        movaps nb134_rinvOO(%rsp),%xmm0
        movss %xmm0,%xmm1
        mulss  nb134_rsqOO(%rsp),%xmm1   ## xmm1=r 
        mulss  nb134_tsc(%rsp),%xmm1

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

    movq nb134_VFtab(%rbp),%rsi
    movd %mm6,%r8d

    ## dispersion 
    movlps (%rsi,%r8,4),%xmm5
    movaps %xmm5,%xmm4
    shufps $136,%xmm7,%xmm4 ## constant 10001000
    shufps $221,%xmm7,%xmm5 ## constant 11011101

    movlps 8(%rsi,%r8,4),%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  nb134_two(%rsp),%xmm7         ## two*Heps2 
    addss  %xmm6,%xmm7
    addss  %xmm5,%xmm7 ## xmm7=FF 
    mulss  %xmm1,%xmm5 ## xmm5=eps*Fp 
    addss  %xmm4,%xmm5 ## xmm5=VV 

    movss nb134_c6(%rsp),%xmm4
    mulss  %xmm4,%xmm7   ## fijD 
    mulss  %xmm4,%xmm5   ## Vvdw6 
        xorps  %xmm3,%xmm3

        mulps  nb134_tsc(%rsp),%xmm7
        subss  %xmm7,%xmm3
        movss  %xmm3,nb134_fstmp(%rsp)

    addss  nb134_Vvdwtot(%rsp),%xmm5
    movss %xmm5,nb134_Vvdwtot(%rsp)

    ## repulsion 
    movlps 16(%rsi,%r8,4),%xmm5
    movaps %xmm5,%xmm4
    shufps $136,%xmm7,%xmm4 ## constant 10001000
    shufps $221,%xmm7,%xmm5 ## constant 11011101

    movlps 24(%rsi,%r8,4),%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  nb134_two(%rsp),%xmm7         ## two*Heps2 
    addss  %xmm6,%xmm7
    addss  %xmm5,%xmm7 ## xmm7=FF 
    mulss  %xmm1,%xmm5 ## xmm5=eps*Fp 
    addss  %xmm4,%xmm5 ## xmm5=VV 

    movss nb134_c12(%rsp),%xmm4
    mulss  %xmm4,%xmm7 ## fijR 
    mulss  %xmm4,%xmm5 ## Vvdw12 
        movaps nb134_fstmp(%rsp),%xmm3
        mulss  nb134_tsc(%rsp),%xmm7
        subss  %xmm7,%xmm3

    addss  nb134_Vvdwtot(%rsp),%xmm5
    movss %xmm5,nb134_Vvdwtot(%rsp)

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

        mulss  nb134_dxOO(%rsp),%xmm0
        mulss  nb134_dyOO(%rsp),%xmm1
        mulss  nb134_dzOO(%rsp),%xmm2
        xorps   %xmm3,%xmm3
        xorps   %xmm4,%xmm4
        xorps   %xmm5,%xmm5
        addss   %xmm0,%xmm3
        addss   %xmm1,%xmm4
        addss   %xmm2,%xmm5
        movaps  %xmm3,nb134_fjxO(%rsp)
        movaps  %xmm4,nb134_fjyO(%rsp)
        movaps  %xmm5,nb134_fjzO(%rsp)
        addss   nb134_fixO(%rsp),%xmm0
        addss   nb134_fiyO(%rsp),%xmm1
        addss   nb134_fizO(%rsp),%xmm2
        movss  %xmm0,nb134_fixO(%rsp)
        movss  %xmm1,nb134_fiyO(%rsp)
        movss  %xmm2,nb134_fizO(%rsp)

        ## do  M coulomb interaction
        movaps nb134_rinvMM(%rsp),%xmm0
        movaps %xmm0,%xmm7      ## xmm7=rinv 
        mulps  %xmm0,%xmm0      ## xmm0=rinvsq 

        ## fetch charges to xmm3 (temporary) 
        xorps  %xmm3,%xmm3
        movss   nb134_qqMH(%rsp),%xmm3
        movhps  nb134_qqMM(%rsp),%xmm3
        shufps $193,%xmm3,%xmm3 ## constant 11000001 

        mulps  %xmm3,%xmm7 ## vcoul=rinv*qq
        movaps %xmm7,%xmm6
        addps  nb134_vctot(%rsp),%xmm7
        movaps %xmm7,nb134_vctot(%rsp)
        mulps  %xmm6,%xmm0

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

        mulps   nb134_dxMM(%rsp),%xmm0
        mulps   nb134_dyMM(%rsp),%xmm1
        mulps   nb134_dzMM(%rsp),%xmm2
        ## update forces M - j water 
        movaps  nb134_fjxO(%rsp),%xmm3
        movaps  nb134_fjyO(%rsp),%xmm4
        movaps  nb134_fjzO(%rsp),%xmm5
        addps   %xmm0,%xmm3
        addps   %xmm1,%xmm4
        addps   %xmm2,%xmm5
        movaps  %xmm3,nb134_fjxO(%rsp)
        movaps  %xmm4,nb134_fjyO(%rsp)
        movaps  %xmm5,nb134_fjzO(%rsp)
        addps   nb134_fixM(%rsp),%xmm0
        addps   nb134_fiyM(%rsp),%xmm1
        addps   nb134_fizM(%rsp),%xmm2
        movaps  %xmm0,nb134_fixM(%rsp)
        movaps  %xmm1,nb134_fiyM(%rsp)
        movaps  %xmm2,nb134_fizM(%rsp)

        ## i H1 & H2 simultaneously first get i particle coords: 
    movaps  nb134_jxO(%rsp),%xmm0
    movaps  nb134_jyO(%rsp),%xmm1
    movaps  nb134_jzO(%rsp),%xmm2
    movaps  %xmm0,%xmm3
    movaps  %xmm1,%xmm4
    movaps  %xmm2,%xmm5
        subps   nb134_ixH1(%rsp),%xmm0
        subps   nb134_iyH1(%rsp),%xmm1
        subps   nb134_izH1(%rsp),%xmm2
        subps   nb134_ixH2(%rsp),%xmm3
        subps   nb134_iyH2(%rsp),%xmm4
        subps   nb134_izH2(%rsp),%xmm5
        movaps %xmm0,nb134_dxH1H1(%rsp)
        movaps %xmm1,nb134_dyH1H1(%rsp)
        movaps %xmm2,nb134_dzH1H1(%rsp)
        movaps %xmm3,nb134_dxH2H2(%rsp)
        movaps %xmm4,nb134_dyH2H2(%rsp)
        movaps %xmm5,nb134_dzH2H2(%rsp)
        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,nb134_rsqH1H1(%rsp)
        movaps  %xmm4,nb134_rsqH2H2(%rsp)

        ## 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  nb134_three(%rsp),%xmm3
        movaps  %xmm3,%xmm7
        mulps   %xmm0,%xmm1
        mulps   %xmm4,%xmm5
        subps   %xmm1,%xmm3
        subps   %xmm5,%xmm7
        mulps   %xmm2,%xmm3
        mulps   %xmm6,%xmm7
        mulps   nb134_half(%rsp),%xmm3   ## rinvH1H1
        mulps   nb134_half(%rsp),%xmm7   ## rinvH2H2
        movaps  %xmm3,nb134_rinvH1H1(%rsp)
        movaps  %xmm7,nb134_rinvH2H2(%rsp)

        ## Do H1 coulomb interaction
        movaps %xmm3,%xmm0
        movaps %xmm0,%xmm7      ## xmm7=rinv 
        mulps  %xmm0,%xmm0      ## xmm0=rinvsq 

        ## fetch charges to xmm3 (temporary) 
        xorps  %xmm3,%xmm3
        movss   nb134_qqHH(%rsp),%xmm3
        movhps  nb134_qqMH(%rsp),%xmm3
        shufps $193,%xmm3,%xmm3 ## constant 11000001 

        mulps  %xmm3,%xmm7 ##vcoul 
        mulps  %xmm7,%xmm0

        addps  nb134_vctot(%rsp),%xmm7
        movaps %xmm7,nb134_vctot(%rsp)

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

        mulps   nb134_dxH1H1(%rsp),%xmm0
        mulps   nb134_dyH1H1(%rsp),%xmm1
        mulps   nb134_dzH1H1(%rsp),%xmm2
        ## update forces H1 - j water 
        movaps  nb134_fjxO(%rsp),%xmm3
        movaps  nb134_fjyO(%rsp),%xmm4
        movaps  nb134_fjzO(%rsp),%xmm5
        addps   %xmm0,%xmm3
        addps   %xmm1,%xmm4
        addps   %xmm2,%xmm5
        movaps  %xmm3,nb134_fjxO(%rsp)
        movaps  %xmm4,nb134_fjyO(%rsp)
        movaps  %xmm5,nb134_fjzO(%rsp)
        addps   nb134_fixH1(%rsp),%xmm0
        addps   nb134_fiyH1(%rsp),%xmm1
        addps   nb134_fizH1(%rsp),%xmm2
        movaps  %xmm0,nb134_fixH1(%rsp)
        movaps  %xmm1,nb134_fiyH1(%rsp)
        movaps  %xmm2,nb134_fizH1(%rsp)

        ## H2 Coulomb
        movaps nb134_rinvH2H2(%rsp),%xmm0
        movaps %xmm0,%xmm7      ## xmm7=rinv 
        mulps  %xmm0,%xmm0      ## xmm0=rinvsq 

        ## fetch charges to xmm3 (temporary) 
        xorps  %xmm3,%xmm3
        movss   nb134_qqHH(%rsp),%xmm3
        movhps  nb134_qqMH(%rsp),%xmm3
        shufps $193,%xmm3,%xmm3 ## constant 11000001 

        mulps  %xmm3,%xmm7
        mulps  %xmm7,%xmm0

        addps  nb134_vctot(%rsp),%xmm7   ## local vctot summation variable
        movaps %xmm7,nb134_vctot(%rsp)

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

        mulps   nb134_dxH2H2(%rsp),%xmm0
        mulps   nb134_dyH2H2(%rsp),%xmm1
        mulps   nb134_dzH2H2(%rsp),%xmm2
        ## update forces H2 - j water 
        movaps  nb134_fjxO(%rsp),%xmm3
        movaps  nb134_fjyO(%rsp),%xmm4
        movaps  nb134_fjzO(%rsp),%xmm5
        addps   %xmm0,%xmm3
        addps   %xmm1,%xmm4
        addps   %xmm2,%xmm5
        movaps  %xmm3,nb134_fjxO(%rsp)
        movaps  %xmm4,nb134_fjyO(%rsp)
        movaps  %xmm5,nb134_fjzO(%rsp)
        addps   nb134_fixH2(%rsp),%xmm0
        addps   nb134_fiyH2(%rsp),%xmm1
        addps   nb134_fizH2(%rsp),%xmm2
        movaps  %xmm0,nb134_fixH2(%rsp)
        movaps  %xmm1,nb134_fiyH2(%rsp)
        movaps  %xmm2,nb134_fizH2(%rsp)

        movq    nb134_faction(%rbp),%rsi
        ## update j water forces from local variables.
        ## transpose back first
        movaps  nb134_fjxO(%rsp),%xmm0   ## Ox H1x H2x Mx 
        movaps  nb134_fjyO(%rsp),%xmm1   ## Oy H1y H2y My
        movaps  nb134_fjzO(%rsp),%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  $50,%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  (%rsi,%rax,4),%xmm0
        movlps  16(%rsi,%rax,4),%xmm1
        movlps  32(%rsi,%rax,4),%xmm4
        movhps  8(%rsi,%rax,4),%xmm0
        movhps  24(%rsi,%rax,4),%xmm1
        movhps  40(%rsi,%rax,4),%xmm4
        addps   %xmm3,%xmm0
        addps   %xmm5,%xmm1
        addps   %xmm2,%xmm4
        movlps   %xmm0,(%rsi,%rax,4)
        movlps   %xmm1,16(%rsi,%rax,4)
        movlps   %xmm4,32(%rsi,%rax,4)
        movhps   %xmm0,8(%rsi,%rax,4)
        movhps   %xmm1,24(%rsi,%rax,4)
        movhps   %xmm4,40(%rsi,%rax,4)

        decl nb134_innerk(%rsp)
        jz    _nb_kernel134_x86_64_sse.nb134_updateouterdata
        jmp   _nb_kernel134_x86_64_sse.nb134_single_loop
_nb_kernel134_x86_64_sse.nb134_updateouterdata: 
        movl  nb134_ii3(%rsp),%ecx
        movq  nb134_faction(%rbp),%rdi
        movq  nb134_fshift(%rbp),%rsi
        movl  nb134_is3(%rsp),%edx

        ## accumulate  Oi forces in xmm0, xmm1, xmm2 
        movaps nb134_fixO(%rsp),%xmm0
        movaps nb134_fiyO(%rsp),%xmm1
        movaps nb134_fizO(%rsp),%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  (%rdi,%rcx,4),%xmm3
        movss  4(%rdi,%rcx,4),%xmm4
        movss  8(%rdi,%rcx,4),%xmm5
        subss  %xmm0,%xmm3
        subss  %xmm1,%xmm4
        subss  %xmm2,%xmm5
        movss  %xmm3,(%rdi,%rcx,4)
        movss  %xmm4,4(%rdi,%rcx,4)
        movss  %xmm5,8(%rdi,%rcx,4)

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