nb_kernel134_ia32_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,068 行 · 第 1/5 页

    mulps  nb134_two(%esp),%xmm7         ## two*Heps2 
    addps  %xmm6,%xmm7
    addps  %xmm5,%xmm7 ## xmm7=FF 
    mulps  %xmm1,%xmm5 ## xmm5=eps*Fp 
    addps  %xmm4,%xmm5 ## xmm5=VV 

    movaps nb134_c12(%esp),%xmm4
    mulps  %xmm4,%xmm7 ## fijR 
    mulps  %xmm4,%xmm5 ## Vvdw12 
    addps  nb134_fstmp(%esp),%xmm7
    mulps nb134_tsc(%esp),%xmm7

    addps  nb134_Vvdwtot(%esp),%xmm5
    movaps %xmm5,nb134_Vvdwtot(%esp)

        xorps %xmm2,%xmm2
        mulps  %xmm0,%xmm7
        subps %xmm7,%xmm2

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

    movd  %mm0,%eax
    movd  %mm1,%ebx
    movd  %mm2,%ecx
    movd  %mm3,%edx

        xorps %xmm3,%xmm3
        movaps %xmm3,%xmm4
        movaps %xmm3,%xmm5
        mulps nb134_dxOO(%esp),%xmm0
        mulps nb134_dyOO(%esp),%xmm1
        mulps nb134_dzOO(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb134_fixO(%esp),%xmm0
        addps nb134_fiyO(%esp),%xmm1
        addps nb134_fizO(%esp),%xmm2
        movaps %xmm3,nb134_fjxO(%esp)
        movaps %xmm4,nb134_fjyO(%esp)
        movaps %xmm5,nb134_fjzO(%esp)
        movaps %xmm0,nb134_fixO(%esp)
        movaps %xmm1,nb134_fiyO(%esp)
        movaps %xmm2,nb134_fizO(%esp)

        ## Coulomb interactions 
        ## start with H1-H1 interaction 
        movaps  nb134_rinvH1H1(%esp),%xmm7
        movaps  %xmm7,%xmm0
        mulps   %xmm0,%xmm0     ## xmm7=rinv, xmm0=rinvsq
        mulps  nb134_qqHH(%esp),%xmm7
        mulps  %xmm7,%xmm0      ## total fs H1 in xmm0 

        movaps %xmm7,%xmm6 ## local vctot summation variable

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

        xorps %xmm3,%xmm3
        movaps %xmm3,%xmm4
        movaps %xmm3,%xmm5
        mulps nb134_dxH1H1(%esp),%xmm0
        mulps nb134_dyH1H1(%esp),%xmm1
        mulps nb134_dzH1H1(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb134_fixH1(%esp),%xmm0
        addps nb134_fiyH1(%esp),%xmm1
        addps nb134_fizH1(%esp),%xmm2
        movaps %xmm3,nb134_fjxH1(%esp)
        movaps %xmm4,nb134_fjyH1(%esp)
        movaps %xmm5,nb134_fjzH1(%esp)
        movaps %xmm0,nb134_fixH1(%esp)
        movaps %xmm1,nb134_fiyH1(%esp)
        movaps %xmm2,nb134_fizH1(%esp)

        ## H1-H2 interaction 
        movaps  nb134_rinvH1H2(%esp),%xmm7
        movaps  %xmm7,%xmm0
        mulps   %xmm0,%xmm0     ## xmm7=rinv, xmm0=rinvsq
        mulps  nb134_qqHH(%esp),%xmm7
        mulps  %xmm7,%xmm0      ## total fs H1 in xmm0 

        addps  %xmm7,%xmm6 ## local vctot summation variable

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

        xorps %xmm3,%xmm3
        movaps %xmm3,%xmm4
        movaps %xmm3,%xmm5
        mulps nb134_dxH1H2(%esp),%xmm0
        mulps nb134_dyH1H2(%esp),%xmm1
        mulps nb134_dzH1H2(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb134_fixH1(%esp),%xmm0
        addps nb134_fiyH1(%esp),%xmm1
        addps nb134_fizH1(%esp),%xmm2
        movaps %xmm3,nb134_fjxH2(%esp)
        movaps %xmm4,nb134_fjyH2(%esp)
        movaps %xmm5,nb134_fjzH2(%esp)
        movaps %xmm0,nb134_fixH1(%esp)
        movaps %xmm1,nb134_fiyH1(%esp)
        movaps %xmm2,nb134_fizH1(%esp)

        ## H1-M interaction  
        movaps  nb134_rinvH1M(%esp),%xmm7
        movaps  %xmm7,%xmm0
        mulps   %xmm0,%xmm0     ## xmm7=rinv, xmm0=rinvsq
        mulps  nb134_qqMH(%esp),%xmm7
        mulps  %xmm7,%xmm0      ## total fs H1 in xmm0 

        addps  %xmm7,%xmm6 ## local vctot summation variable

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

        xorps %xmm3,%xmm3
        movaps %xmm3,%xmm4
        movaps %xmm3,%xmm5
        mulps nb134_dxH1M(%esp),%xmm0
        mulps nb134_dyH1M(%esp),%xmm1
        mulps nb134_dzH1M(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb134_fixH1(%esp),%xmm0
        addps nb134_fiyH1(%esp),%xmm1
        addps nb134_fizH1(%esp),%xmm2
        movaps %xmm3,nb134_fjxM(%esp)
        movaps %xmm4,nb134_fjyM(%esp)
        movaps %xmm5,nb134_fjzM(%esp)
        movaps %xmm0,nb134_fixH1(%esp)
        movaps %xmm1,nb134_fiyH1(%esp)
        movaps %xmm2,nb134_fizH1(%esp)

        ## H2-H1 interaction 
        movaps  nb134_rinvH2H1(%esp),%xmm7
        movaps  %xmm7,%xmm0
        mulps   %xmm0,%xmm0     ## xmm7=rinv, xmm0=rinvsq
        mulps  nb134_qqHH(%esp),%xmm7
        mulps  %xmm7,%xmm0      ## total fs H1 in xmm0 

        addps  %xmm7,%xmm6 ## local vctot summation variable

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

        movaps nb134_fjxH1(%esp),%xmm3
        movaps nb134_fjyH1(%esp),%xmm4
        movaps nb134_fjzH1(%esp),%xmm5
        mulps nb134_dxH2H1(%esp),%xmm0
        mulps nb134_dyH2H1(%esp),%xmm1
        mulps nb134_dzH2H1(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb134_fixH2(%esp),%xmm0
        addps nb134_fiyH2(%esp),%xmm1
        addps nb134_fizH2(%esp),%xmm2
        movaps %xmm3,nb134_fjxH1(%esp)
        movaps %xmm4,nb134_fjyH1(%esp)
        movaps %xmm5,nb134_fjzH1(%esp)
        movaps %xmm0,nb134_fixH2(%esp)
        movaps %xmm1,nb134_fiyH2(%esp)
        movaps %xmm2,nb134_fizH2(%esp)

        ## H2-H2 interaction 
        movaps  nb134_rinvH2H2(%esp),%xmm7
        movaps  %xmm7,%xmm0
        mulps   %xmm0,%xmm0     ## xmm7=rinv, xmm0=rinvsq
        mulps  nb134_qqHH(%esp),%xmm7
        mulps  %xmm7,%xmm0      ## total fs H1 in xmm0 

        addps  %xmm7,%xmm6 ## local vctot summation variable

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

        movaps nb134_fjxH2(%esp),%xmm3
        movaps nb134_fjyH2(%esp),%xmm4
        movaps nb134_fjzH2(%esp),%xmm5
        mulps nb134_dxH2H2(%esp),%xmm0
        mulps nb134_dyH2H2(%esp),%xmm1
        mulps nb134_dzH2H2(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb134_fixH2(%esp),%xmm0
        addps nb134_fiyH2(%esp),%xmm1
        addps nb134_fizH2(%esp),%xmm2
        movaps %xmm3,nb134_fjxH2(%esp)
        movaps %xmm4,nb134_fjyH2(%esp)
        movaps %xmm5,nb134_fjzH2(%esp)
        movaps %xmm0,nb134_fixH2(%esp)
        movaps %xmm1,nb134_fiyH2(%esp)
        movaps %xmm2,nb134_fizH2(%esp)

        ## H2-M interaction 
        movaps  nb134_rinvH2M(%esp),%xmm7
        movaps  %xmm7,%xmm0
        mulps   %xmm0,%xmm0     ## xmm7=rinv, xmm0=rinvsq
        mulps  nb134_qqMH(%esp),%xmm7
        mulps  %xmm7,%xmm0      ## total fs H1 in xmm0 

        addps  %xmm7,%xmm6 ## local vctot summation variable

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

        movaps nb134_fjxM(%esp),%xmm3
        movaps nb134_fjyM(%esp),%xmm4
        movaps nb134_fjzM(%esp),%xmm5
        mulps nb134_dxH2M(%esp),%xmm0
        mulps nb134_dyH2M(%esp),%xmm1
        mulps nb134_dzH2M(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb134_fixH2(%esp),%xmm0
        addps nb134_fiyH2(%esp),%xmm1
        addps nb134_fizH2(%esp),%xmm2
        movaps %xmm3,nb134_fjxM(%esp)
        movaps %xmm4,nb134_fjyM(%esp)
        movaps %xmm5,nb134_fjzM(%esp)
        movaps %xmm0,nb134_fixH2(%esp)
        movaps %xmm1,nb134_fiyH2(%esp)
        movaps %xmm2,nb134_fizH2(%esp)

        ## M-H1 interaction 
        movaps  nb134_rinvMH1(%esp),%xmm7
        movaps  %xmm7,%xmm0
        mulps   %xmm0,%xmm0     ## xmm7=rinv, xmm0=rinvsq
        mulps  nb134_qqMH(%esp),%xmm7
        mulps  %xmm7,%xmm0      ## total fs in xmm0 

        addps  %xmm7,%xmm6 ## local vctot summation variable

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

        movaps nb134_fjxH1(%esp),%xmm3
        movaps nb134_fjyH1(%esp),%xmm4
        movaps nb134_fjzH1(%esp),%xmm5
        mulps nb134_dxMH1(%esp),%xmm0
        mulps nb134_dyMH1(%esp),%xmm1
        mulps nb134_dzMH1(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb134_fixM(%esp),%xmm0
        addps nb134_fiyM(%esp),%xmm1
        addps nb134_fizM(%esp),%xmm2
        movaps %xmm3,nb134_fjxH1(%esp)
        movaps %xmm4,nb134_fjyH1(%esp)
        movaps %xmm5,nb134_fjzH1(%esp)
        movaps %xmm0,nb134_fixM(%esp)
        movaps %xmm1,nb134_fiyM(%esp)
        movaps %xmm2,nb134_fizM(%esp)

        ## M-H2 interaction 
        movaps  nb134_rinvMH2(%esp),%xmm7
        movaps  %xmm7,%xmm0
        mulps   %xmm0,%xmm0     ## xmm7=rinv, xmm0=rinvsq
        mulps  nb134_qqMH(%esp),%xmm7
        mulps  %xmm7,%xmm0      ## total fs in xmm0 

        addps  %xmm7,%xmm6 ## local vctot summation variable

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

        movaps nb134_fjxH2(%esp),%xmm3
        movaps nb134_fjyH2(%esp),%xmm4
        movaps nb134_fjzH2(%esp),%xmm5
        mulps nb134_dxMH2(%esp),%xmm0
        mulps nb134_dyMH2(%esp),%xmm1
        mulps nb134_dzMH2(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb134_fixM(%esp),%xmm0
        addps nb134_fiyM(%esp),%xmm1
        addps nb134_fizM(%esp),%xmm2
        movaps %xmm3,nb134_fjxH2(%esp)
        movaps %xmm4,nb134_fjyH2(%esp)
        movaps %xmm5,nb134_fjzH2(%esp)
        movaps %xmm0,nb134_fixM(%esp)
        movaps %xmm1,nb134_fiyM(%esp)
        movaps %xmm2,nb134_fizM(%esp)

        ## M-M interaction 
        movaps  nb134_rinvMM(%esp),%xmm7
        movaps  %xmm7,%xmm0
        mulps   %xmm0,%xmm0     ## xmm7=rinv, xmm0=rinvsq
        mulps  nb134_qqMM(%esp),%xmm7
        mulps  %xmm7,%xmm0      ## total fs in xmm0 

        addps  %xmm7,%xmm6 ## local vctot summation variable

        movaps %xmm0,%xmm1
        addps  nb134_vctot(%esp),%xmm6
        movaps %xmm6,nb134_vctot(%esp)
        movaps %xmm0,%xmm2

        movaps nb134_fjxM(%esp),%xmm3
        movaps nb134_fjyM(%esp),%xmm4
        movaps nb134_fjzM(%esp),%xmm5
        mulps nb134_dxMM(%esp),%xmm0
        mulps nb134_dyMM(%esp),%xmm1
        mulps nb134_dzMM(%esp),%xmm2
        subps %xmm0,%xmm3
        subps %xmm1,%xmm4
        subps %xmm2,%xmm5
        addps nb134_fixM(%esp),%xmm0
        addps nb134_fiyM(%esp),%xmm1
        addps nb134_fizM(%esp),%xmm2
        movaps %xmm3,nb134_fjxM(%esp)
        movaps %xmm4,nb134_fjyM(%esp)
        movaps %xmm5,nb134_fjzM(%esp)
        movaps %xmm0,nb134_fixM(%esp)
        movaps %xmm1,nb134_fiyM(%esp)
        movaps %xmm2,nb134_fizM(%esp)

        movl nb134_faction(%ebp),%edi
        ## update j forces 
        ## 4 j waters with four atoms each.
        ## step 1 : transpose fjxO, fjyO, fjzO, fjxH1
        movaps nb134_fjxO(%esp),%xmm0
        movaps nb134_fjyO(%esp),%xmm1
        movaps nb134_fjzO(%esp),%xmm2
        movaps nb134_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 nb134_fjyH1(%esp),%xmm0
        movaps nb134_fjzH1(%esp),%xmm1
        movaps nb134_fjxH2(%esp),%xmm2
        movaps nb134_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)