nb_kernel332_x86_64_sse2.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,381 行 · 第 1/5 页

    movapd %xmm8,%xmm7
    movapd %xmm8,%xmm9
    movapd %xmm11,%xmm10
    movapd %xmm11,%xmm12

        mulpd nb332_dxOH2(%rsp),%xmm3
        mulpd nb332_dyOH2(%rsp),%xmm4
        mulpd nb332_dzOH2(%rsp),%xmm5
        mulpd nb332_dxH1H2(%rsp),%xmm7
        mulpd nb332_dyH1H2(%rsp),%xmm8
        mulpd nb332_dzH1H2(%rsp),%xmm9
        mulpd nb332_dxH2H2(%rsp),%xmm10
        mulpd nb332_dyH2H2(%rsp),%xmm11
        mulpd nb332_dzH2H2(%rsp),%xmm12

    addpd %xmm3,%xmm0
    addpd %xmm4,%xmm1
    addpd %xmm5,%xmm2
    addpd nb332_fixO(%rsp),%xmm3
    addpd nb332_fiyO(%rsp),%xmm4
    addpd nb332_fizO(%rsp),%xmm5

    addpd %xmm7,%xmm0
    addpd %xmm8,%xmm1
    addpd %xmm9,%xmm2
    addpd nb332_fixH1(%rsp),%xmm7
    addpd nb332_fiyH1(%rsp),%xmm8
    addpd nb332_fizH1(%rsp),%xmm9

    addpd %xmm10,%xmm0
    addpd %xmm11,%xmm1
    addpd %xmm12,%xmm2
    addpd nb332_fixH2(%rsp),%xmm10
    addpd nb332_fiyH2(%rsp),%xmm11
    addpd nb332_fizH2(%rsp),%xmm12

    movapd %xmm3,nb332_fixO(%rsp)
    movapd %xmm4,nb332_fiyO(%rsp)
    movapd %xmm5,nb332_fizO(%rsp)
    movapd %xmm7,nb332_fixH1(%rsp)
    movapd %xmm8,nb332_fiyH1(%rsp)
    movapd %xmm9,nb332_fizH1(%rsp)
    movapd %xmm10,nb332_fixH2(%rsp)
    movapd %xmm11,nb332_fiyH2(%rsp)
    movapd %xmm12,nb332_fizH2(%rsp)

    ## store back j H2 forces from xmm0-xmm2
        movlpd %xmm0,48(%rdi,%rax,8)
        movlpd %xmm1,56(%rdi,%rax,8)
        movlpd %xmm2,64(%rdi,%rax,8)
        movhpd %xmm0,48(%rdi,%rbx,8)
        movhpd %xmm1,56(%rdi,%rbx,8)
        movhpd %xmm2,64(%rdi,%rbx,8)

        ## should we do one more iteration? 
        subl $2,nb332_innerk(%rsp)
        jl    _nb_kernel332_x86_64_sse2.nb332_checksingle
        jmp   _nb_kernel332_x86_64_sse2.nb332_unroll_loop
_nb_kernel332_x86_64_sse2.nb332_checksingle: 
        movl  nb332_innerk(%rsp),%edx
        andl  $1,%edx
        jnz   _nb_kernel332_x86_64_sse2.nb332_dosingle
        jmp   _nb_kernel332_x86_64_sse2.nb332_updateouterdata
_nb_kernel332_x86_64_sse2.nb332_dosingle: 
        movq  nb332_innerjjnr(%rsp),%rdx       ## pointer to jjnr[k] 
        movl  (%rdx),%eax

        movq nb332_pos(%rbp),%rsi
        lea  (%rax,%rax,2),%rax


        ## move j O coordinates to local temp variables 
    movsd (%rsi,%rax,8),%xmm0
    movsd 8(%rsi,%rax,8),%xmm1
    movsd 16(%rsi,%rax,8),%xmm2

    ## xmm0 = Ox
    ## xmm1 = Oy
    ## xmm2 = Oz

    movsd %xmm0,%xmm3
    movsd %xmm1,%xmm4
    movsd %xmm2,%xmm5
    movsd %xmm0,%xmm6
    movsd %xmm1,%xmm7
    movsd %xmm2,%xmm8

    subsd nb332_ixO(%rsp),%xmm0
    subsd nb332_iyO(%rsp),%xmm1
    subsd nb332_izO(%rsp),%xmm2
    subsd nb332_ixH1(%rsp),%xmm3
    subsd nb332_iyH1(%rsp),%xmm4
    subsd nb332_izH1(%rsp),%xmm5
    subsd nb332_ixH2(%rsp),%xmm6
    subsd nb332_iyH2(%rsp),%xmm7
    subsd nb332_izH2(%rsp),%xmm8

        movsd %xmm0,nb332_dxOO(%rsp)
        movsd %xmm1,nb332_dyOO(%rsp)
        movsd %xmm2,nb332_dzOO(%rsp)
        mulsd  %xmm0,%xmm0
        mulsd  %xmm1,%xmm1
        mulsd  %xmm2,%xmm2
        movsd %xmm3,nb332_dxH1O(%rsp)
        movsd %xmm4,nb332_dyH1O(%rsp)
        movsd %xmm5,nb332_dzH1O(%rsp)
        mulsd  %xmm3,%xmm3
        mulsd  %xmm4,%xmm4
        mulsd  %xmm5,%xmm5
        movsd %xmm6,nb332_dxH2O(%rsp)
        movsd %xmm7,nb332_dyH2O(%rsp)
        movsd %xmm8,nb332_dzH2O(%rsp)
        mulsd  %xmm6,%xmm6
        mulsd  %xmm7,%xmm7
        mulsd  %xmm8,%xmm8
        addsd  %xmm1,%xmm0
        addsd  %xmm2,%xmm0
        addsd  %xmm4,%xmm3
        addsd  %xmm5,%xmm3
    addsd  %xmm7,%xmm6
    addsd  %xmm8,%xmm6

        ## start doing invsqrt for jO atoms
    cvtsd2ss %xmm0,%xmm1
    cvtsd2ss %xmm3,%xmm4
    cvtsd2ss %xmm6,%xmm7
        rsqrtss %xmm1,%xmm1
        rsqrtss %xmm4,%xmm4
    rsqrtss %xmm7,%xmm7
    cvtss2sd %xmm1,%xmm1
    cvtss2sd %xmm4,%xmm4
    cvtss2sd %xmm7,%xmm7

        movsd  %xmm1,%xmm2
        movsd  %xmm4,%xmm5
    movsd  %xmm7,%xmm8

        mulsd   %xmm1,%xmm1 ## lu*lu
        mulsd   %xmm4,%xmm4 ## lu*lu
    mulsd   %xmm7,%xmm7 ## lu*lu

        movsd  nb332_three(%rsp),%xmm9
        movsd  %xmm9,%xmm10
    movsd  %xmm9,%xmm11

        mulsd   %xmm0,%xmm1 ## rsq*lu*lu
        mulsd   %xmm3,%xmm4 ## rsq*lu*lu 
    mulsd   %xmm6,%xmm7 ## rsq*lu*lu

        subsd   %xmm1,%xmm9
        subsd   %xmm4,%xmm10
    subsd   %xmm7,%xmm11 ## 3-rsq*lu*lu

        mulsd   %xmm2,%xmm9
        mulsd   %xmm5,%xmm10
    mulsd   %xmm8,%xmm11 ## lu*(3-rsq*lu*lu)

        movsd  nb332_half(%rsp),%xmm15
        mulsd   %xmm15,%xmm9 ## first iteration for rinvOO 
        mulsd   %xmm15,%xmm10 ## first iteration for rinvH1O
    mulsd   %xmm15,%xmm11 ## first iteration for rinvH2O 

    ## second iteration step    
        movsd  %xmm9,%xmm2
        movsd  %xmm10,%xmm5
    movsd  %xmm11,%xmm8

        mulsd   %xmm2,%xmm2 ## lu*lu
        mulsd   %xmm5,%xmm5 ## lu*lu
    mulsd   %xmm8,%xmm8 ## lu*lu

        movsd  nb332_three(%rsp),%xmm1
        movsd  %xmm1,%xmm4
    movsd  %xmm1,%xmm7

        mulsd   %xmm0,%xmm2 ## rsq*lu*lu
        mulsd   %xmm3,%xmm5 ## rsq*lu*lu 
    mulsd   %xmm6,%xmm8 ## rsq*lu*lu

        subsd   %xmm2,%xmm1
        subsd   %xmm5,%xmm4
    subsd   %xmm8,%xmm7 ## 3-rsq*lu*lu

        mulsd   %xmm1,%xmm9
        mulsd   %xmm4,%xmm10
    mulsd   %xmm7,%xmm11 ## lu*(3-rsq*lu*lu)

        movsd  nb332_half(%rsp),%xmm15
        mulsd   %xmm15,%xmm9 ##  rinvOO 
        mulsd   %xmm15,%xmm10 ##   rinvH1O
    mulsd   %xmm15,%xmm11 ##   rinvH2O

        movsd  %xmm9,nb332_rinvOO(%rsp)
        movsd  %xmm10,nb332_rinvH1O(%rsp)
        movsd  %xmm11,nb332_rinvH2O(%rsp)

        ## O interactions 
    ## rsq in xmm0,xmm3,xmm6  
    ## rinv in xmm9, xmm10, xmm11

    movapd %xmm9,nb332_rinvOO(%rsp)
    movsd  nb332_tsc(%rsp),%xmm1

    mulsd  %xmm9,%xmm0 ## r
    mulsd  %xmm10,%xmm3
    mulsd  %xmm11,%xmm6
    mulsd  %xmm1,%xmm0 ## rtab
    mulsd  %xmm1,%xmm3
    mulsd  %xmm1,%xmm6

    ## truncate and convert to integers
    cvttsd2si %xmm0,%r8d
    cvttsd2si %xmm3,%r10d
    cvttsd2si %xmm6,%r12d

    ## convert back to float
    cvtsi2sd  %r8d,%xmm2
    cvtsi2sd  %r10d,%xmm5
    cvtsi2sd  %r12d,%xmm8

    ## multiply by 4
    shlq  $2,%r8
    shll  $2,%r10d
    shll  $2,%r12d

        ## multiply by 3
    lea  (%r8,%r8,2),%r8
    lea  (%r10,%r10,2),%r10
    lea  (%r12,%r12,2),%r12

    movq nb332_VFtab(%rbp),%rsi

    ## calculate eps
    subsd     %xmm2,%xmm0
    subsd     %xmm5,%xmm3
    subsd     %xmm8,%xmm6

    movsd    %xmm0,nb332_epsO(%rsp)
    movsd    %xmm3,nb332_epsH1(%rsp)
    movsd    %xmm6,nb332_epsH2(%rsp)

    ## Load LOTS of table data
    movsd (%rsi,%r8,8),%xmm0
    movsd 8(%rsi,%r8,8),%xmm1
    movsd 16(%rsi,%r8,8),%xmm2
    movsd 24(%rsi,%r8,8),%xmm3
    movsd (%rsi,%r10,8),%xmm4
    movsd 8(%rsi,%r10,8),%xmm5
    movsd 16(%rsi,%r10,8),%xmm6
    movsd 24(%rsi,%r10,8),%xmm7
    movsd (%rsi,%r12,8),%xmm8
    movsd 8(%rsi,%r12,8),%xmm9
    movsd 16(%rsi,%r12,8),%xmm10
    movsd 24(%rsi,%r12,8),%xmm11
    ## table data ready in xmm0-xmm3 , xmm4-xmm7 , and xmm8-xmm11

    movsd nb332_epsO(%rsp),%xmm12
    movsd nb332_epsH1(%rsp),%xmm13
    movsd nb332_epsH2(%rsp),%xmm14

    mulsd  %xmm12,%xmm3  ## Heps
    mulsd  %xmm13,%xmm7
    mulsd  %xmm14,%xmm11
    mulsd  %xmm12,%xmm2  ## Geps
    mulsd  %xmm13,%xmm6
    mulsd  %xmm14,%xmm10
    mulsd  %xmm12,%xmm3  ## Heps2
    mulsd  %xmm13,%xmm7
    mulsd  %xmm14,%xmm11

    addsd  %xmm2,%xmm1  ## F+Geps
    addsd  %xmm6,%xmm5
    addsd  %xmm10,%xmm9
    addsd  %xmm3,%xmm1  ## F+Geps+Heps2 = Fp
    addsd  %xmm7,%xmm5
    addsd  %xmm11,%xmm9
    addsd  %xmm3,%xmm3   ## 2*Heps2
    addsd  %xmm7,%xmm7
    addsd  %xmm11,%xmm11
    addsd  %xmm2,%xmm3   ## 2*Heps2+Geps
    addsd  %xmm6,%xmm7
    addsd  %xmm10,%xmm11
    addsd  %xmm1,%xmm3  ## FF = Fp + 2*Heps2 + Geps
    addsd  %xmm5,%xmm7
    addsd  %xmm9,%xmm11
    mulsd  nb332_epsO(%rsp),%xmm1     ## eps*Fp
    mulsd  nb332_epsH1(%rsp),%xmm5
    mulsd  nb332_epsH2(%rsp),%xmm9
    movsd nb332_qqOO(%rsp),%xmm12
    movsd nb332_qqOH(%rsp),%xmm13
    addsd  %xmm0,%xmm1    ## VV
    addsd  %xmm4,%xmm5
    addsd  %xmm8,%xmm9
    mulsd  nb332_qqOO(%rsp),%xmm1     ## VV*qq = vcoul
    mulsd  nb332_qqOH(%rsp),%xmm5
    mulsd  nb332_qqOH(%rsp),%xmm9
    mulsd  nb332_qqOO(%rsp),%xmm3      ## FF*qq = fij
    mulsd  nb332_qqOH(%rsp),%xmm7
    mulsd  nb332_qqOH(%rsp),%xmm11

    ## accumulate vctot
    addsd  nb332_vctot(%rsp),%xmm1
    addsd  %xmm9,%xmm5
    addsd  %xmm5,%xmm1
    movsd %xmm1,nb332_vctot(%rsp)

    movsd %xmm7,%xmm2
    movsd %xmm11,%xmm1

    ## fij coul in xmm3, xmm2, xmm1    

    ## calculate LJ table
    movsd 32(%rsi,%r8,8),%xmm4
    movsd 40(%rsi,%r8,8),%xmm5
    movsd 48(%rsi,%r8,8),%xmm6
    movsd 56(%rsi,%r8,8),%xmm7
    movsd 64(%rsi,%r8,8),%xmm8
    movsd 72(%rsi,%r8,8),%xmm9
    movsd 80(%rsi,%r8,8),%xmm10
    movsd 88(%rsi,%r8,8),%xmm11
    ## dispersion table in xmm4-xmm7, repulsion table in xmm8-xmm11

    movsd nb332_epsO(%rsp),%xmm0

    mulsd  %xmm0,%xmm7   ## Heps
    mulsd  %xmm0,%xmm11
    mulsd  %xmm0,%xmm6  ## Geps
    mulsd  %xmm0,%xmm10
    mulsd  %xmm0,%xmm7  ## Heps2
    mulsd  %xmm0,%xmm11
    addsd  %xmm6,%xmm5 ## F+Geps
    addsd  %xmm10,%xmm9
    addsd  %xmm7,%xmm5  ## F+Geps+Heps2 = Fp
    addsd  %xmm11,%xmm9
    addsd  %xmm7,%xmm7   ## 2*Heps2
    addsd  %xmm11,%xmm11
    addsd  %xmm6,%xmm7  ## 2*Heps2+Geps
    addsd  %xmm10,%xmm11

    addsd  %xmm5,%xmm7 ## FF = Fp + 2*Heps2 + Geps
    addsd  %xmm9,%xmm11
    mulsd  %xmm0,%xmm5 ## eps*Fp
    mulsd  %xmm0,%xmm9
    movsd nb332_c6(%rsp),%xmm12
    movsd nb332_c12(%rsp),%xmm13
    addsd  %xmm4,%xmm5 ## VV
    addsd  %xmm8,%xmm9

    mulsd  %xmm12,%xmm5 ## VV*c6 = vnb6
    mulsd  %xmm13,%xmm9 ## VV*c12 = vnb12
    addsd  %xmm9,%xmm5
    addsd  nb332_Vvdwtot(%rsp),%xmm5
    movsd %xmm5,nb332_Vvdwtot(%rsp)

    mulsd  %xmm12,%xmm7  ## FF*c6 = fnb6
    mulsd  %xmm13,%xmm11  ## FF*c12  = fnb12
    addsd  %xmm11,%xmm7

    addsd  %xmm7,%xmm3
    movsd nb332_tsc(%rsp),%xmm10

    mulsd  %xmm10,%xmm3 ## fscal
    mulsd  %xmm10,%xmm2
    mulsd  %xmm10,%xmm1

    ## move j O forces to xmm11-xmm13
    movq nb332_faction(%rbp),%rdi
        movsd (%rdi,%rax,8),%xmm11
        movsd 8(%rdi,%rax,8),%xmm12
        movsd 16(%rdi,%rax,8),%xmm13

    xorpd  %xmm0,%xmm0
    xorpd  %xmm4,%xmm4
    xorpd  %xmm8,%xmm8

    subsd  %xmm3,%xmm0
    subsd  %xmm2,%xmm4
    subsd  %xmm1,%xmm8

    mulsd  nb332_rinvOO(%rsp),%xmm0
    mulsd  nb332_rinvH1O(%rsp),%xmm4
    mulsd  nb332_rinvH2O(%rsp),%xmm8

    movsd %xmm0,%xmm1
    movsd %xmm0,%xmm2
    movsd %xmm4,%xmm3
    movsd %xmm4,%xmm5
    movsd %xmm8,%xmm6
    movsd %xmm8,%xmm7

        mulsd nb332_dxOO(%rsp),%xmm0
        mulsd nb332_dyOO(%rsp),%xmm1
        mulsd nb332_dzOO(%rsp),%xmm2
        mulsd nb332_dxH1O(%rsp),%xmm3
        mulsd nb332_dyH1O(%rsp),%xmm4
        mulsd nb332_dzH1O(%rsp),%xmm5
        mulsd nb332_dxH2O(%rsp),%xmm6
        mulsd nb332_dyH2O(%rsp),%xmm7
        mulsd nb332_dzH2O(%rsp),%xmm8

    addsd %xmm0,%xmm11
    addsd %xmm1,%xmm12
    addsd %xmm2,%xmm13
    addsd nb332_fixO(%rsp),%xmm0
    addsd nb332_fiyO(%rsp),%xmm1
    addsd nb332_fizO(%rsp),%xmm2

    addsd %xmm3,%xmm11
    addsd %xmm4,%xmm12
    addsd %xmm5,%xmm13
    addsd nb332_fixH1(%rsp),%xmm3
    addsd nb332_fiyH1(%rsp),%xmm4
    addsd nb332_fizH1(%rsp),%xmm5

    addsd %xmm6,%xmm11
    addsd %xmm7,%xmm12
    addsd %xmm8,%xmm13
    addsd nb332_fixH2(%rsp),%xmm6
    addsd nb332_fiyH2(%rsp),%xmm7
    addsd nb332_fizH2(%rsp),%xmm8

    movsd %xmm0,nb332_fixO(%rsp)
    movsd %xmm1,nb332_fiyO(%rsp)
    movsd %xmm2,nb332_fizO(%rsp)
    movsd %xmm3,nb332_fixH1(%rsp)
    movsd %xmm4,nb332_fiyH1(%rsp)
    movsd %xmm5,nb332_fizH1(%rsp)
    movsd %xmm6,nb332_fixH2(%rsp)
    movsd %xmm7,nb332_fiyH2(%rsp)
    movsd %xmm8,nb332_fizH2(%rsp)

    ## store back j O forces from xmm11-xmm13
        movsd %xmm11,(%rdi,%rax,8)
        movsd %xmm12,8(%rdi,%rax,8)
        movsd %xmm13,16(%rdi,%rax,8)

        ## move j H1 coordinates to local temp variables 
    movq nb332_pos(%rbp),%rsi
    movsd 24(%rsi,%rax,8),%xmm0
    movsd 32(%rsi,%rax,8),%xmm1
    movsd 40(%rsi,%rax,8),%xmm2

    ## xmm0 = H1x
    ## xmm1 = H1y
    ## xmm2 = H1z

    movsd %xmm0,%xmm3
    movsd %xmm1,%xmm4
    movsd %xmm2,%xmm5
    movsd %xmm0,%xmm6
    movsd %xmm1,%xmm7
    movsd %xmm2,%xmm8

    subsd nb332_ixO(%rsp),%xmm0
    subsd nb332_iyO(%rsp),%xmm1
    subsd nb332_izO(%rsp),%xmm2
    subsd nb332_ixH1(%rsp),%xmm3
    subsd nb332_iyH1(%rsp),%xmm4
    subsd nb332_izH1(%rsp),%xmm5
    subsd nb332_ixH2(%rsp),%xmm6
    subsd nb332_iyH2(%rsp),%xmm7
    subsd nb332_izH2(%rsp),%xmm8

        movsd %xmm0,nb332_dxOH1(%rsp)
        movsd %xmm1,nb332_dyOH1(%rsp)
        movsd %xmm2,nb332_dzOH1(%rsp)
        mulsd  %xmm0,%xmm0
        mulsd  %xmm1,%xmm1
        mulsd  %xmm2,%xmm2
        movsd %xmm3,nb332_dxH1H1(%rsp)
        movsd %xmm4,nb332_dyH1H1(%rsp)
        movsd %xmm5,nb332_dzH1H1(%rsp)
        mulsd  %xmm3,%xmm3
        mulsd  %xmm4,%xmm4
        mulsd  %xmm5,%xmm5
        movsd %xmm6,nb332_dxH2H1(%rsp)
        movsd %xmm7,nb332_dyH2H1(%rsp)