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

        mulps nb302_tsc(%rsp),%xmm3
        mulps %xmm0,%xmm3
        subps  %xmm3,%xmm1

        movaps  %xmm1,%xmm0
        movaps  %xmm1,%xmm2
        mulps   nb302_dxH1O(%rsp),%xmm0
        mulps   nb302_dyH1O(%rsp),%xmm1
        mulps   nb302_dzH1O(%rsp),%xmm2
        ## update forces H1 - j water 
        movaps  nb302_fjxO(%rsp),%xmm3
        movaps  nb302_fjyO(%rsp),%xmm4
        movaps  nb302_fjzO(%rsp),%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_fixH1(%rsp),%xmm0
        addps   nb302_fiyH1(%rsp),%xmm1
        addps   nb302_fizH1(%rsp),%xmm2
        movaps  %xmm0,nb302_fixH1(%rsp)
        movaps  %xmm1,nb302_fiyH1(%rsp)
        movaps  %xmm2,nb302_fizH1(%rsp)
        ## do table for H2 - j water interaction 
        movaps nb302_rinvH2O(%rsp),%xmm0
        movaps nb302_rsqH2O(%rsp),%xmm1
        mulps  %xmm0,%xmm1      ## xmm0=rinv, xmm1=r 
        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

        mulps nb302_tsc(%rsp),%xmm3
        mulps %xmm0,%xmm3
        subps  %xmm3,%xmm1

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

        mulps   nb302_dxH2O(%rsp),%xmm0
        mulps   nb302_dyH2O(%rsp),%xmm1
        mulps   nb302_dzH2O(%rsp),%xmm2
        movaps  nb302_fjxO(%rsp),%xmm3
        movaps  nb302_fjyO(%rsp),%xmm4
        movaps  nb302_fjzO(%rsp),%xmm5
        addps   %xmm0,%xmm3
        addps   %xmm1,%xmm4
        addps   %xmm2,%xmm5
        movq    nb302_faction(%rbp),%rsi
        movaps  %xmm3,nb302_fjxO(%rsp)
        movaps  %xmm4,nb302_fjyO(%rsp)
        movaps  %xmm5,nb302_fjzO(%rsp)
        addps   nb302_fixH2(%rsp),%xmm0
        addps   nb302_fiyH2(%rsp),%xmm1
        addps   nb302_fizH2(%rsp),%xmm2
        movaps  %xmm0,nb302_fixH2(%rsp)
        movaps  %xmm1,nb302_fiyH2(%rsp)
        movaps  %xmm2,nb302_fizH2(%rsp)

        ## update j water forces from local variables 
        movlps  (%rsi,%rax,4),%xmm0
        movlps  12(%rsi,%rax,4),%xmm1
        movhps  24(%rsi,%rax,4),%xmm1
        movaps  nb302_fjxO(%rsp),%xmm3
        movaps  nb302_fjyO(%rsp),%xmm4
        movaps  nb302_fjzO(%rsp),%xmm5
        movaps  %xmm5,%xmm6
        movaps  %xmm5,%xmm7
        shufps $2,%xmm6,%xmm6 ## 00000010
        shufps $3,%xmm7,%xmm7 ## 00000011
        addss   8(%rsi,%rax,4),%xmm5
        addss   20(%rsi,%rax,4),%xmm6
        addss   32(%rsi,%rax,4),%xmm7
        movss   %xmm5,8(%rsi,%rax,4)
        movss   %xmm6,20(%rsi,%rax,4)
        movss   %xmm7,32(%rsi,%rax,4)
        movaps   %xmm3,%xmm5
        unpcklps %xmm4,%xmm3
        unpckhps %xmm4,%xmm5
        addps    %xmm3,%xmm0
        addps    %xmm5,%xmm1
        movlps  %xmm0,(%rsi,%rax,4)
        movlps  %xmm1,12(%rsi,%rax,4)
        movhps  %xmm1,24(%rsi,%rax,4)

        decl nb302_innerk(%rsp)
        jz    _nb_kernel302_x86_64_sse.nb302_updateouterdata
        jmp   _nb_kernel302_x86_64_sse.nb302_single_loop
_nb_kernel302_x86_64_sse.nb302_updateouterdata: 
        movl  nb302_ii3(%rsp),%ecx
        movq  nb302_faction(%rbp),%rdi
        movq  nb302_fshift(%rbp),%rsi
        movl  nb302_is3(%rsp),%edx

        ## accumulate  Oi forces in xmm0, xmm1, xmm2 
        movaps nb302_fixO(%rsp),%xmm0
        movaps nb302_fiyO(%rsp),%xmm1
        movaps nb302_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 ## 00001000       

        ## accumulate H1i forces in xmm0, xmm1, xmm2 
        movaps nb302_fixH1(%rsp),%xmm0
        movaps nb302_fiyH1(%rsp),%xmm1
        movaps nb302_fizH1(%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  12(%rdi,%rcx,4),%xmm3
        movss  16(%rdi,%rcx,4),%xmm4
        movss  20(%rdi,%rcx,4),%xmm5
        subss  %xmm0,%xmm3
        subss  %xmm1,%xmm4
        subss  %xmm2,%xmm5
        movss  %xmm3,12(%rdi,%rcx,4)
        movss  %xmm4,16(%rdi,%rcx,4)
        movss  %xmm5,20(%rdi,%rcx,4)

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

        ## accumulate H2i forces in xmm0, xmm1, xmm2 
        movaps nb302_fixH2(%rsp),%xmm0
        movaps nb302_fiyH2(%rsp),%xmm1
        movaps nb302_fizH2(%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  24(%rdi,%rcx,4),%xmm3
        movss  28(%rdi,%rcx,4),%xmm4
        movss  32(%rdi,%rcx,4),%xmm5
        subss  %xmm0,%xmm3
        subss  %xmm1,%xmm4
        subss  %xmm2,%xmm5
        movss  %xmm3,24(%rdi,%rcx,4)
        movss  %xmm4,28(%rdi,%rcx,4)
        movss  %xmm5,32(%rdi,%rcx,4)

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

        ## increment fshift force  
        movlps  (%rsi,%rdx,4),%xmm3
        movss  8(%rsi,%rdx,4),%xmm4
        subps  %xmm6,%xmm3
        subss  %xmm7,%xmm4
        movlps  %xmm3,(%rsi,%rdx,4)
        movss  %xmm4,8(%rsi,%rdx,4)

        ## get n from stack
        movl nb302_n(%rsp),%esi
        ## get group index for i particle 
        movq  nb302_gid(%rbp),%rdx              ## base of gid[]
        movl  (%rdx,%rsi,4),%edx                ## ggid=gid[n]

        ## accumulate total potential energy and update it 
        movaps nb302_vctot(%rsp),%xmm7
        ## accumulate 
        movhlps %xmm7,%xmm6
        addps  %xmm6,%xmm7      ## pos 0-1 in xmm7 have the sum now 
        movaps %xmm7,%xmm6
        shufps $1,%xmm6,%xmm6
        addss  %xmm6,%xmm7

        ## add earlier value from mem 
        movq  nb302_Vc(%rbp),%rax
        addss (%rax,%rdx,4),%xmm7
        ## move back to mem 
        movss %xmm7,(%rax,%rdx,4)

        ## finish if last 
        movl nb302_nn1(%rsp),%ecx
        ## esi already loaded with n
        incl %esi
        subl %esi,%ecx
        jz _nb_kernel302_x86_64_sse.nb302_outerend

        ## not last, iterate outer loop once more!  
        movl %esi,nb302_n(%rsp)
        jmp _nb_kernel302_x86_64_sse.nb302_outer
_nb_kernel302_x86_64_sse.nb302_outerend: 
        ## check if more outer neighborlists remain
        movl  nb302_nri(%rsp),%ecx
        ## esi already loaded with n above
        subl  %esi,%ecx
        jz _nb_kernel302_x86_64_sse.nb302_end
        ## non-zero, do one more workunit
        jmp   _nb_kernel302_x86_64_sse.nb302_threadloop
_nb_kernel302_x86_64_sse.nb302_end: 

        movl nb302_nouter(%rsp),%eax
        movl nb302_ninner(%rsp),%ebx
        movq nb302_outeriter(%rbp),%rcx
        movq nb302_inneriter(%rbp),%rdx
        movl %eax,(%rcx)
        movl %ebx,(%rdx)

        addq $1528,%rsp
        emms

        pop %r15
        pop %r14
        pop %r13
        pop %r12

        pop %rbx
        pop    %rbp
        ret






.globl nb_kernel302nf_x86_64_sse
.globl _nb_kernel302nf_x86_64_sse
nb_kernel302nf_x86_64_sse:      
_nb_kernel302nf_x86_64_sse:     
##      Room for return address and rbp (16 bytes)
.set nb302nf_fshift, 16
.set nb302nf_gid, 24
.set nb302nf_pos, 32
.set nb302nf_faction, 40
.set nb302nf_charge, 48
.set nb302nf_p_facel, 56
.set nb302nf_argkrf, 64
.set nb302nf_argcrf, 72
.set nb302nf_Vc, 80
.set nb302nf_type, 88
.set nb302nf_p_ntype, 96
.set nb302nf_vdwparam, 104
.set nb302nf_Vvdw, 112
.set nb302nf_p_tabscale, 120
.set nb302nf_VFtab, 128
.set nb302nf_invsqrta, 136
.set nb302nf_dvda, 144
.set nb302nf_p_gbtabscale, 152
.set nb302nf_GBtab, 160
.set nb302nf_p_nthreads, 168
.set nb302nf_count, 176
.set nb302nf_mtx, 184
.set nb302nf_outeriter, 192
.set nb302nf_inneriter, 200
.set nb302nf_work, 208
        ## stack offsets for local variables  
        ## bottom of stack is cache-aligned for sse use 
.set nb302nf_ixO, 0
.set nb302nf_iyO, 16
.set nb302nf_izO, 32
.set nb302nf_ixH1, 48
.set nb302nf_iyH1, 64
.set nb302nf_izH1, 80
.set nb302nf_ixH2, 96
.set nb302nf_iyH2, 112
.set nb302nf_izH2, 128
.set nb302nf_jxO, 144
.set nb302nf_jyO, 160
.set nb302nf_jzO, 176
.set nb302nf_jxH1, 192
.set nb302nf_jyH1, 208
.set nb302nf_jzH1, 224
.set nb302nf_jxH2, 240
.set nb302nf_jyH2, 256
.set nb302nf_jzH2, 272
.set nb302nf_qqOO, 288
.set nb302nf_qqOH, 304
.set nb302nf_qqHH, 320
.set nb302nf_tsc, 336
.set nb302nf_vctot, 352
.set nb302nf_half, 368
.set nb302nf_three, 384
.set nb302nf_rsqOO, 400
.set nb302nf_rsqOH1, 416
.set nb302nf_rsqOH2, 432
.set nb302nf_rsqH1O, 448
.set nb302nf_rsqH1H1, 464
.set nb302nf_rsqH1H2, 480
.set nb302nf_rsqH2O, 496
.set nb302nf_rsqH2H1, 512
.set nb302nf_rsqH2H2, 528
.set nb302nf_rinvOO, 544
.set nb302nf_rinvOH1, 560
.set nb302nf_rinvOH2, 576
.set nb302nf_rinvH1O, 592
.set nb302nf_rinvH1H1, 608
.set nb302nf_rinvH1H2, 624
.set nb302nf_rinvH2O, 640
.set nb302nf_rinvH2H1, 656
.set nb302nf_rinvH2H2, 672
.set nb302nf_is3, 688
.set nb302nf_ii3, 692
.set nb302nf_nri, 696
.set nb302nf_iinr, 704
.set nb302nf_jindex, 712
.set nb302nf_jjnr, 720
.set nb302nf_shift, 728
.set nb302nf_shiftvec, 736
.set nb302nf_facel, 744
.set nb302nf_innerjjnr, 752
.set nb302nf_innerk, 760
.set nb302nf_n, 764
.set nb302nf_nn1, 768
.set nb302nf_nouter, 772
.set nb302nf_ninner, 776
        push %rbp
        movq %rsp,%rbp
        push %rbx

        emms
        subq $792,%rsp          ## local variable stack space (n*4+8)

        ## zero 32-bit iteration counters
        movl $0,%eax
        movl %eax,nb302nf_nouter(%rsp)
        movl %eax,nb302nf_ninner(%rsp)

        movl (%rdi),%edi
        movl %edi,nb302nf_nri(%rsp)
        movq %rsi,nb302nf_iinr(%rsp)
        movq %rdx,nb302nf_jindex(%rsp)
        movq %rcx,nb302nf_jjnr(%rsp)