nb_kernel302_x86_64_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,240 行 · 第 1/5 页

        movhps 8(%rsi,%r9,4),%xmm7
        movhps 8(%rsi,%r13,4),%xmm11

        movhps 8(%rsi,%rdx,4),%xmm12
        movhps 8(%rsi,%r11,4),%xmm13
        movhps 8(%rsi,%r15,4),%xmm14

    movaps %xmm3,%xmm2
    movaps %xmm7,%xmm6
    movaps %xmm11,%xmm10

        shufps $136,%xmm12,%xmm2 ## 10001000
        shufps $136,%xmm13,%xmm6 ## 10001000
        shufps $136,%xmm14,%xmm10 ## 10001000
        shufps $221,%xmm12,%xmm3 ## 11011101
        shufps $221,%xmm13,%xmm7 ## 11011101
        shufps $221,%xmm14,%xmm11 ## 11011101
    ## table data ready in xmm0-xmm3 , xmm4-xmm7 , and xmm8-xmm11

    movaps nb302_epsO(%rsp),%xmm12
    movaps nb302_epsH1(%rsp),%xmm13
    movaps nb302_epsH2(%rsp),%xmm14

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

    addps  %xmm2,%xmm1  ## F+Geps
    addps  %xmm6,%xmm5
    addps  %xmm10,%xmm9
    addps  %xmm3,%xmm1  ## F+Geps+Heps2 = Fp
    addps  %xmm7,%xmm5
    addps  %xmm11,%xmm9
    addps  %xmm3,%xmm3   ## 2*Heps2
    addps  %xmm7,%xmm7
    addps  %xmm11,%xmm11
    addps  %xmm2,%xmm3   ## 2*Heps2+Geps
    addps  %xmm6,%xmm7
    addps  %xmm10,%xmm11
    addps  %xmm1,%xmm3  ## FF = Fp + 2*Heps2 + Geps
    addps  %xmm5,%xmm7
    addps  %xmm9,%xmm11
    mulps  %xmm12,%xmm1  ## eps*Fp
    mulps  %xmm13,%xmm5
    mulps  %xmm14,%xmm9
    movaps nb302_qqOH(%rsp),%xmm12
    movaps nb302_qqHH(%rsp),%xmm13
    addps  %xmm0,%xmm1    ## VV
    addps  %xmm4,%xmm5
    addps  %xmm8,%xmm9
    mulps  %xmm12,%xmm1  ## VV*qq = vcoul
    mulps  %xmm13,%xmm5
    mulps  %xmm13,%xmm9
    mulps  %xmm12,%xmm3   ## FF*qq = fij
    mulps  %xmm13,%xmm7
    mulps  %xmm13,%xmm11

    ## accumulate vctot
    addps  nb302_vctot(%rsp),%xmm1
    addps  %xmm9,%xmm5
    addps  %xmm5,%xmm1
    movaps %xmm1,nb302_vctot(%rsp)

    movaps nb302_tsc(%rsp),%xmm10
    mulps  %xmm10,%xmm3 ## fscal
    mulps  %xmm10,%xmm7
    mulps  %xmm11,%xmm10

    movd %mm0,%eax ## restore j3 from mm0-mm3
    movd %mm1,%ebx
    movd %mm2,%ecx
    movd %mm3,%edx

        ## move j H2 forces to local temp variables 
    movlps 24(%rdi,%rax,4),%xmm11    ## jxH2a jyH2a  -   -
    movlps 24(%rdi,%rcx,4),%xmm12    ## jxH2c jyH2c  -   -
    movhps 24(%rdi,%rbx,4),%xmm11    ## jxH2a jyH2a jxH2b jyH2b 
    movhps 24(%rdi,%rdx,4),%xmm12    ## jxH2c jyH2c jxH2d jyH2d 

    movss  32(%rdi,%rax,4),%xmm13    ## jzH2a  -  -  -
    movss  32(%rdi,%rcx,4),%xmm14    ## jzH2c  -  -  -
    movss  32(%rdi,%rbx,4),%xmm1    ## jzH2b  -  -  -
    movss  32(%rdi,%rdx,4),%xmm2    ## jzH2d  -  -  -
    movlhps %xmm1,%xmm13 ## jzH2a  -  jzH2b  -
    movlhps %xmm2,%xmm14 ## jzH2c  -  jzH2d -

    shufps $136,%xmm14,%xmm13 ## 10001000 => jzH2a jzH2b jzH2c jzH2d

    ## xmm11: jxH2a jyH2a jxH2b jyH2b 
    ## xmm12: jxH2c jyH2c jxH2d jyH2d
    ## xmm13: jzH2a jzH2b jzH2c jzH2d

    xorps  %xmm0,%xmm0
    xorps  %xmm4,%xmm4
    xorps  %xmm8,%xmm8

    mulps  nb302_rinvOH2(%rsp),%xmm3
    mulps  nb302_rinvH1H2(%rsp),%xmm7
    mulps  nb302_rinvH2H2(%rsp),%xmm10

    subps  %xmm3,%xmm0
    subps  %xmm7,%xmm4
    subps  %xmm10,%xmm8

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

        mulps nb302_dxOH2(%rsp),%xmm0
        mulps nb302_dyOH2(%rsp),%xmm1
        mulps nb302_dzOH2(%rsp),%xmm2
        mulps nb302_dxH1H2(%rsp),%xmm3
        mulps nb302_dyH1H2(%rsp),%xmm4
        mulps nb302_dzH1H2(%rsp),%xmm5
        mulps nb302_dxH2H2(%rsp),%xmm6
        mulps nb302_dyH2H2(%rsp),%xmm7
        mulps nb302_dzH2H2(%rsp),%xmm8

    movaps %xmm0,%xmm14
    movaps %xmm1,%xmm15
    addps %xmm2,%xmm13
    addps nb302_fixO(%rsp),%xmm0
    addps nb302_fiyO(%rsp),%xmm1
    addps nb302_fizO(%rsp),%xmm2

    addps %xmm3,%xmm14
    addps %xmm4,%xmm15
    addps %xmm5,%xmm13
    addps nb302_fixH1(%rsp),%xmm3
    addps nb302_fiyH1(%rsp),%xmm4
    addps nb302_fizH1(%rsp),%xmm5

    addps %xmm6,%xmm14
    addps %xmm7,%xmm15
    addps %xmm8,%xmm13
    addps nb302_fixH2(%rsp),%xmm6
    addps nb302_fiyH2(%rsp),%xmm7
    addps nb302_fizH2(%rsp),%xmm8

    movaps %xmm0,nb302_fixO(%rsp)
    movaps %xmm1,nb302_fiyO(%rsp)
    movaps %xmm2,nb302_fizO(%rsp)
    movaps %xmm3,nb302_fixH1(%rsp)
    movaps %xmm4,nb302_fiyH1(%rsp)
    movaps %xmm5,nb302_fizH1(%rsp)
    movaps %xmm6,nb302_fixH2(%rsp)
    movaps %xmm7,nb302_fiyH2(%rsp)
    movaps %xmm8,nb302_fizH2(%rsp)

    ## xmm14 = fH2x
    ## xmm15 = fH2y
    ## xmm13 = fH2z
    movaps %xmm14,%xmm0
    unpcklps %xmm15,%xmm14
    unpckhps %xmm15,%xmm0

    addps  %xmm14,%xmm11
    addps  %xmm0,%xmm12

    movhlps  %xmm13,%xmm14 ## fH2zc fH2zd

    movlps %xmm11,24(%rdi,%rax,4)
    movhps %xmm11,24(%rdi,%rbx,4)
    movlps %xmm12,24(%rdi,%rcx,4)
    movhps %xmm12,24(%rdi,%rdx,4)
    movss  %xmm13,32(%rdi,%rax,4)
    movss  %xmm14,32(%rdi,%rcx,4)
    shufps $1,%xmm13,%xmm13
    shufps $1,%xmm14,%xmm14
    movss  %xmm13,32(%rdi,%rbx,4)
    movss  %xmm14,32(%rdi,%rdx,4)

        ## should we do one more iteration? 
        subl $4,nb302_innerk(%rsp)
        jl    _nb_kernel302_x86_64_sse.nb302_single_check
        jmp   _nb_kernel302_x86_64_sse.nb302_unroll_loop
_nb_kernel302_x86_64_sse.nb302_single_check: 
        addl $4,nb302_innerk(%rsp)
        jnz   _nb_kernel302_x86_64_sse.nb302_single_loop
        jmp   _nb_kernel302_x86_64_sse.nb302_updateouterdata
_nb_kernel302_x86_64_sse.nb302_single_loop: 
        movq  nb302_innerjjnr(%rsp),%rdx        ## pointer to jjnr[k] 
        movl  (%rdx),%eax
        addq $4,nb302_innerjjnr(%rsp)

        movq nb302_pos(%rbp),%rsi
        lea  (%rax,%rax,2),%rax

        ## fetch j coordinates 
        xorps %xmm0,%xmm0
        xorps %xmm1,%xmm1
        xorps %xmm2,%xmm2

        movss (%rsi,%rax,4),%xmm0               ## jxO  -  -  -
        movss 4(%rsi,%rax,4),%xmm1              ## jyO  -  -  -
        movss 8(%rsi,%rax,4),%xmm2              ## jzO  -  -  -  

        movlps 12(%rsi,%rax,4),%xmm6            ## xmm6 = jxH1 jyH1   -    -
        movss  20(%rsi,%rax,4),%xmm7            ## xmm7 = jzH1   -    -    - 
        movhps 24(%rsi,%rax,4),%xmm6            ## xmm6 = jxH1 jyH1 jxH2 jyH2
        movss  32(%rsi,%rax,4),%xmm5            ## xmm5 = jzH2   -    -    -

        ## have all coords, time for some shuffling.

        shufps $216,%xmm6,%xmm6 ## 11011000      ;# xmm6 = jxH1 jxH2 jyH1 jyH2 
        unpcklps %xmm5,%xmm7                    ## xmm7 = jzH1 jzH2   -    -

        movlhps %xmm6,%xmm0                     ## xmm0 = jxO   0   jxH1 jxH2 
        shufps $228,%xmm6,%xmm1 ## 11100100     ;# xmm1 = jyO   0   jyH1 jyH2 
        shufps $68,%xmm7,%xmm2 ## 01000100     ;# xmm2 = jzO   0   jzH1 jzH2

        ## store all j coordinates in jO  
        movaps %xmm0,nb302_jxO(%rsp)
        movaps %xmm1,nb302_jyO(%rsp)
        movaps %xmm2,nb302_jzO(%rsp)
        subps  nb302_ixO(%rsp),%xmm0
        subps  nb302_iyO(%rsp),%xmm1
        subps  nb302_izO(%rsp),%xmm2
        movaps %xmm0,nb302_dxOO(%rsp)
        movaps %xmm1,nb302_dyOO(%rsp)
        movaps %xmm2,nb302_dzOO(%rsp)
        mulps %xmm0,%xmm0
        mulps %xmm1,%xmm1
        mulps %xmm2,%xmm2
        addps %xmm1,%xmm0
        addps %xmm2,%xmm0       ## have rsq in xmm0 

        ## do invsqrt 
        rsqrtps %xmm0,%xmm1
        movaps  %xmm1,%xmm2
        mulps   %xmm1,%xmm1
        movaps  nb302_three(%rsp),%xmm3
        mulps   %xmm0,%xmm1
        subps   %xmm1,%xmm3
        mulps   %xmm2,%xmm3
        mulps   nb302_half(%rsp),%xmm3   ## rinv iO - j water 

        movaps  %xmm3,%xmm1
        mulps   %xmm0,%xmm1     ## xmm1=r 
        movaps  %xmm3,%xmm0     ## xmm0=rinv 
        mulps  nb302_tsc(%rsp),%xmm1

        movhlps %xmm1,%xmm2
        cvttps2pi %xmm1,%mm6
        cvttps2pi %xmm2,%mm7    ## mm6/mm7 contain lu indices 
        cvtpi2ps %mm6,%xmm3
        cvtpi2ps %mm7,%xmm2
        movlhps  %xmm2,%xmm3
        subps    %xmm3,%xmm1    ## xmm1=eps 
        movaps %xmm1,%xmm2
        mulps  %xmm2,%xmm2      ## xmm2=eps2 
        pslld   $2,%mm6
        pslld   $2,%mm7

        movd %mm6,%ebx
        movd %mm7,%ecx
        psrlq $32,%mm7
        movd %mm7,%edx          ## table indices in ebx,ecx,edx 

        movq nb302_VFtab(%rbp),%rsi

        movlps (%rsi,%rbx,4),%xmm5
        movlps (%rsi,%rcx,4),%xmm7
        movhps (%rsi,%rdx,4),%xmm7 ## got half coulomb table 
        movaps %xmm5,%xmm4
        shufps $136,%xmm7,%xmm4 ## 10001000
        shufps $221,%xmm7,%xmm5 ## 11011101

        movlps 8(%rsi,%rbx,4),%xmm7
        movlps 8(%rsi,%rcx,4),%xmm3
        movhps 8(%rsi,%rdx,4),%xmm3    ## other half of coulomb table  
        movaps %xmm7,%xmm6
        shufps $136,%xmm3,%xmm6 ## 10001000
        shufps $221,%xmm3,%xmm7 ## 11011101
        ## coulomb table ready, in xmm4-xmm7  
        mulps  %xmm1,%xmm6      ## xmm6=Geps 
        mulps  %xmm2,%xmm7      ## xmm7=Heps2 
        addps  %xmm6,%xmm5
        addps  %xmm7,%xmm5      ## xmm5=Fp 
        mulps  nb302_two(%rsp),%xmm7            ## two*Heps2 

        xorps  %xmm3,%xmm3
        ## fetch charges to xmm3 (temporary) 
        movss   nb302_qqOO(%rsp),%xmm3
        movhps  nb302_qqOH(%rsp),%xmm3

        addps  %xmm6,%xmm7
        addps  %xmm5,%xmm7 ## xmm7=FF 
        mulps  %xmm1,%xmm5 ## xmm5=eps*Fp 
        addps  %xmm4,%xmm5 ## xmm5=VV 
        mulps  %xmm3,%xmm5 ## vcoul=qq*VV  
        mulps  %xmm7,%xmm3 ## fijC=FF*qq 
        ## at this point xmm5 contains vcoul and xmm3 fijC 

        addps  nb302_vctot(%rsp),%xmm5
        movaps %xmm5,nb302_vctot(%rsp)
        xorps  %xmm2,%xmm2
        mulps  nb302_tsc(%rsp),%xmm3

        subps  %xmm3,%xmm2
        mulps  %xmm2,%xmm0

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

        mulps   nb302_dxOO(%rsp),%xmm0
        mulps   nb302_dyOO(%rsp),%xmm1
        mulps   nb302_dzOO(%rsp),%xmm2
        ## initial update for j forces 
        xorps   %xmm3,%xmm3
        xorps   %xmm4,%xmm4
        xorps   %xmm5,%xmm5
        addps   %xmm0,%xmm3
        addps   %xmm1,%xmm4
        addps   %xmm2,%xmm5
        movaps  %xmm3,nb302_fjxO(%rsp)
        movaps  %xmm4,nb302_fjyO(%rsp)
        movaps  %xmm5,nb302_fjzO(%rsp)
        addps   nb302_fixO(%rsp),%xmm0
        addps   nb302_fiyO(%rsp),%xmm1
        addps   nb302_fizO(%rsp),%xmm2
        movaps  %xmm0,nb302_fixO(%rsp)
        movaps  %xmm1,nb302_fiyO(%rsp)
        movaps  %xmm2,nb302_fizO(%rsp)


        ## done with i O Now do i H1 & H2 simultaneously first get i particle coords: 
    movaps  nb302_jxO(%rsp),%xmm0
    movaps  nb302_jyO(%rsp),%xmm1
    movaps  nb302_jzO(%rsp),%xmm2
    movaps  %xmm0,%xmm3
    movaps  %xmm1,%xmm4
    movaps  %xmm2,%xmm5
        subps  nb302_ixH1(%rsp),%xmm0
        subps  nb302_iyH1(%rsp),%xmm1
        subps  nb302_izH1(%rsp),%xmm2
        subps  nb302_ixH2(%rsp),%xmm3
        subps  nb302_iyH2(%rsp),%xmm4
        subps  nb302_izH2(%rsp),%xmm5
        movaps %xmm0,nb302_dxH1O(%rsp)
        movaps %xmm1,nb302_dyH1O(%rsp)
        movaps %xmm2,nb302_dzH1O(%rsp)
        movaps %xmm3,nb302_dxH2O(%rsp)
        movaps %xmm4,nb302_dyH2O(%rsp)
        movaps %xmm5,nb302_dzH2O(%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 

        ## start with H1, save H2 data 
        movaps %xmm4,nb302_rsqH2O(%rsp)

        ## do invsqrt 
        rsqrtps %xmm0,%xmm1
        rsqrtps %xmm4,%xmm5
        movaps  %xmm1,%xmm2
        movaps  %xmm5,%xmm6
        mulps   %xmm1,%xmm1
        mulps   %xmm5,%xmm5
        movaps  nb302_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   nb302_half(%rsp),%xmm3   ## rinv H1 - j water 
        mulps   nb302_half(%rsp),%xmm7   ## rinv H2 - j water  

        ## start with H1, save H2 data 
        movaps %xmm7,nb302_rinvH2O(%rsp)

        movaps %xmm3,%xmm1
        mulps  %xmm0,%xmm1      ## xmm1=r 
        movaps %xmm3,%xmm0      ## xmm0=rinv 
        mulps  nb302_tsc(%rsp),%xmm1

        movhlps %xmm1,%xmm2
        cvttps2pi %xmm1,%mm6
        cvttps2pi %xmm2,%mm7    ## mm6/mm7 contain lu indices 
        cvtpi2ps %mm6,%xmm3
        cvtpi2ps %mm7,%xmm2
        movlhps  %xmm2,%xmm3
        subps    %xmm3,%xmm1    ## xmm1=eps 
        movaps %xmm1,%xmm2
        mulps  %xmm2,%xmm2      ## xmm2=eps2 
        pslld   $2,%mm6
        pslld   $2,%mm7

        movd %mm6,%ebx
        movd %mm7,%ecx
        psrlq $32,%mm7
        movd %mm7,%edx          ## table indices in ebx,ecx,edx 

        movlps (%rsi,%rbx,4),%xmm5
        movlps (%rsi,%rcx,4),%xmm7
        movhps (%rsi,%rdx,4),%xmm7 ## got half coulomb table 
        movaps %xmm5,%xmm4
        shufps $136,%xmm7,%xmm4 ## 10001000
        shufps $221,%xmm7,%xmm5 ## 11011101

        movlps 8(%rsi,%rbx,4),%xmm7
        movlps 8(%rsi,%rcx,4),%xmm3
        movhps 8(%rsi,%rdx,4),%xmm3    ## other half of coulomb table  
        movaps %xmm7,%xmm6
        shufps $136,%xmm3,%xmm6 ## 10001000
        shufps $221,%xmm3,%xmm7 ## 11011101
        ## coulomb table ready, in xmm4-xmm7  
        mulps  %xmm1,%xmm6      ## xmm6=Geps 
        mulps  %xmm2,%xmm7      ## xmm7=Heps2 
        addps  %xmm6,%xmm5
        addps  %xmm7,%xmm5      ## xmm5=Fp 
        mulps  nb302_two(%rsp),%xmm7            ## two*Heps2 

        xorps  %xmm3,%xmm3
        ## fetch charges to xmm3 (temporary) 
        movss   nb302_qqOH(%rsp),%xmm3
        movhps  nb302_qqHH(%rsp),%xmm3

        addps  %xmm6,%xmm7
        addps  %xmm5,%xmm7 ## xmm7=FF 
        mulps  %xmm1,%xmm5 ## xmm5=eps*Fp 
        addps  %xmm4,%xmm5 ## xmm5=VV 
        mulps  %xmm3,%xmm5 ## vcoul=qq*VV  
        mulps  %xmm7,%xmm3 ## fijC=FF*qq 
        ## at this point xmm5 contains vcoul and xmm3 fijC 
        addps  nb302_vctot(%rsp),%xmm5
        movaps %xmm5,nb302_vctot(%rsp)

    xorps  %xmm1,%xmm1