nb_kernel211_ia32_sse2.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 1,984 行 · 第 1/5 页

    popl %eax
        leave
        ret




.globl nb_kernel211nf_ia32_sse2
.globl _nb_kernel211nf_ia32_sse2
nb_kernel211nf_ia32_sse2:       
_nb_kernel211nf_ia32_sse2:      
.set nb211nf_p_nri, 8
.set nb211nf_iinr, 12
.set nb211nf_jindex, 16
.set nb211nf_jjnr, 20
.set nb211nf_shift, 24
.set nb211nf_shiftvec, 28
.set nb211nf_fshift, 32
.set nb211nf_gid, 36
.set nb211nf_pos, 40
.set nb211nf_faction, 44
.set nb211nf_charge, 48
.set nb211nf_p_facel, 52
.set nb211nf_argkrf, 56
.set nb211nf_argcrf, 60
.set nb211nf_Vc, 64
.set nb211nf_type, 68
.set nb211nf_p_ntype, 72
.set nb211nf_vdwparam, 76
.set nb211nf_Vvdw, 80
.set nb211nf_p_tabscale, 84
.set nb211nf_VFtab, 88
.set nb211nf_invsqrta, 92
.set nb211nf_dvda, 96
.set nb211nf_p_gbtabscale, 100
.set nb211nf_GBtab, 104
.set nb211nf_p_nthreads, 108
.set nb211nf_count, 112
.set nb211nf_mtx, 116
.set nb211nf_outeriter, 120
.set nb211nf_inneriter, 124
.set nb211nf_work, 128
        ## stack offsets for local variables  
        ## bottom of stack is cache-aligned for sse use 
.set nb211nf_ixO, 0
.set nb211nf_iyO, 16
.set nb211nf_izO, 32
.set nb211nf_ixH1, 48
.set nb211nf_iyH1, 64
.set nb211nf_izH1, 80
.set nb211nf_ixH2, 96
.set nb211nf_iyH2, 112
.set nb211nf_izH2, 128
.set nb211nf_iqO, 144
.set nb211nf_iqH, 160
.set nb211nf_qqO, 176
.set nb211nf_qqH, 192
.set nb211nf_c6, 208
.set nb211nf_c12, 224
.set nb211nf_vctot, 240
.set nb211nf_Vvdwtot, 256
.set nb211nf_half, 272
.set nb211nf_three, 288
.set nb211nf_krf, 304
.set nb211nf_crf, 320
.set nb211nf_krsqO, 336
.set nb211nf_krsqH1, 352
.set nb211nf_krsqH2, 368
.set nb211nf_is3, 384
.set nb211nf_ii3, 388
.set nb211nf_ntia, 392
.set nb211nf_innerjjnr, 396
.set nb211nf_innerk, 400
.set nb211nf_n, 404
.set nb211nf_nn1, 408
.set nb211nf_nri, 412
.set nb211nf_nouter, 416
.set nb211nf_ninner, 420
.set nb211nf_salign, 424
        pushl %ebp
        movl %esp,%ebp
    pushl %eax
    pushl %ebx
    pushl %ecx
    pushl %edx
        pushl %esi
        pushl %edi
        subl $428,%esp          ## local stack space 
        movl %esp,%eax
        andl $0xf,%eax
        subl %eax,%esp
        movl %eax,nb211nf_salign(%esp)

        emms

        ## Move args passed by reference to stack
        movl nb211nf_p_nri(%ebp),%ecx
        movl (%ecx),%ecx
        movl %ecx,nb211nf_nri(%esp)

        ## zero iteration counters
        movl $0,%eax
        movl %eax,nb211nf_nouter(%esp)
        movl %eax,nb211nf_ninner(%esp)


        ## create constant floating-point factors on stack
        movl $0x00000000,%eax   ## lower half of double 0.5 IEEE (hex)
        movl $0x3fe00000,%ebx
        movl %eax,nb211nf_half(%esp)
        movl %ebx,nb211nf_half+4(%esp)
        movsd nb211nf_half(%esp),%xmm1
        shufpd $0,%xmm1,%xmm1  ## splat to all elements
        movapd %xmm1,%xmm3
        addpd  %xmm3,%xmm3      ## 1.0
        movapd %xmm3,%xmm2
        addpd  %xmm2,%xmm2      ## 2.0
        addpd  %xmm2,%xmm3      ## 3.0
        movapd %xmm1,nb211nf_half(%esp)
        movapd %xmm3,nb211nf_three(%esp)

        movl nb211nf_argkrf(%ebp),%esi
        movl nb211nf_argcrf(%ebp),%edi
        movsd (%esi),%xmm5
        movsd (%edi),%xmm6
        shufpd $0,%xmm5,%xmm5
        shufpd $0,%xmm6,%xmm6
        movapd %xmm5,nb211nf_krf(%esp)
        movapd %xmm6,nb211nf_crf(%esp)

        ## assume we have at least one i particle - start directly 
        movl  nb211nf_iinr(%ebp),%ecx         ## ecx = pointer into iinr[]      
        movl  (%ecx),%ebx           ## ebx =ii 

        movl  nb211nf_charge(%ebp),%edx
        movsd (%edx,%ebx,8),%xmm3
        movsd 8(%edx,%ebx,8),%xmm4
        movl nb211nf_p_facel(%ebp),%esi
        movsd (%esi),%xmm5
        mulsd  %xmm5,%xmm3
        mulsd  %xmm5,%xmm4

        shufpd $0,%xmm3,%xmm3
        shufpd $0,%xmm4,%xmm4
        movapd %xmm3,nb211nf_iqO(%esp)
        movapd %xmm4,nb211nf_iqH(%esp)

        movl  nb211nf_type(%ebp),%edx
        movl  (%edx,%ebx,4),%ecx
        shll  %ecx
        movl nb211nf_p_ntype(%ebp),%edi
        imull (%edi),%ecx     ## ecx = ntia = 2*ntype*type[ii0] 
        movl  %ecx,nb211nf_ntia(%esp)
_nb_kernel211nf_ia32_sse2.nb211nf_threadloop: 
        movl  nb211nf_count(%ebp),%esi          ## pointer to sync counter
        movl  (%esi),%eax
_nb_kernel211nf_ia32_sse2.nb211nf_spinlock: 
        movl  %eax,%ebx                         ## ebx=*count=nn0
        addl  $1,%ebx                           ## ebx=nn1=nn0+10
        lock 
        cmpxchgl %ebx,(%esi)                    ## write nn1 to *counter,
                                                ## if it hasnt changed.
                                                ## or reread *counter to eax.
        pause                                   ## -> better p4 performance
        jnz _nb_kernel211nf_ia32_sse2.nb211nf_spinlock

        ## if(nn1>nri) nn1=nri
        movl nb211nf_nri(%esp),%ecx
        movl %ecx,%edx
        subl %ebx,%ecx
        cmovlel %edx,%ebx                       ## if(nn1>nri) nn1=nri
        ## Cleared the spinlock if we got here.
        ## eax contains nn0, ebx contains nn1.
        movl %eax,nb211nf_n(%esp)
        movl %ebx,nb211nf_nn1(%esp)
        subl %eax,%ebx                          ## calc number of outer lists
        movl %eax,%esi                          ## copy n to esi
        jg  _nb_kernel211nf_ia32_sse2.nb211nf_outerstart
        jmp _nb_kernel211nf_ia32_sse2.nb211nf_end

_nb_kernel211nf_ia32_sse2.nb211nf_outerstart: 
        ## ebx contains number of outer iterations
        addl nb211nf_nouter(%esp),%ebx
        movl %ebx,nb211nf_nouter(%esp)

_nb_kernel211nf_ia32_sse2.nb211nf_outer: 
        movl  nb211nf_shift(%ebp),%eax        ## eax = pointer into shift[] 
        movl  (%eax,%esi,4),%ebx        ## ebx=shift[n] 

        leal  (%ebx,%ebx,2),%ebx    ## ebx=3*is 
        movl  %ebx,nb211nf_is3(%esp)            ## store is3 

        movl  nb211nf_shiftvec(%ebp),%eax     ## eax = base of shiftvec[] 

        movsd (%eax,%ebx,8),%xmm0
        movsd 8(%eax,%ebx,8),%xmm1
        movsd 16(%eax,%ebx,8),%xmm2

        movl  nb211nf_iinr(%ebp),%ecx         ## ecx = pointer into iinr[]      
        movl  (%ecx,%esi,4),%ebx    ## ebx =ii 

        movapd %xmm0,%xmm3
        movapd %xmm1,%xmm4
        movapd %xmm2,%xmm5

        leal  (%ebx,%ebx,2),%ebx        ## ebx = 3*ii=ii3 
        movl  nb211nf_pos(%ebp),%eax      ## eax = base of pos[]  
        movl  %ebx,nb211nf_ii3(%esp)

        addsd (%eax,%ebx,8),%xmm3
        addsd 8(%eax,%ebx,8),%xmm4
        addsd 16(%eax,%ebx,8),%xmm5
        shufpd $0,%xmm3,%xmm3
        shufpd $0,%xmm4,%xmm4
        shufpd $0,%xmm5,%xmm5
        movapd %xmm3,nb211nf_ixO(%esp)
        movapd %xmm4,nb211nf_iyO(%esp)
        movapd %xmm5,nb211nf_izO(%esp)

        movsd %xmm0,%xmm3
        movsd %xmm1,%xmm4
        movsd %xmm2,%xmm5
        addsd 24(%eax,%ebx,8),%xmm0
        addsd 32(%eax,%ebx,8),%xmm1
        addsd 40(%eax,%ebx,8),%xmm2
        addsd 48(%eax,%ebx,8),%xmm3
        addsd 56(%eax,%ebx,8),%xmm4
        addsd 64(%eax,%ebx,8),%xmm5

        shufpd $0,%xmm0,%xmm0
        shufpd $0,%xmm1,%xmm1
        shufpd $0,%xmm2,%xmm2
        shufpd $0,%xmm3,%xmm3
        shufpd $0,%xmm4,%xmm4
        shufpd $0,%xmm5,%xmm5
        movapd %xmm0,nb211nf_ixH1(%esp)
        movapd %xmm1,nb211nf_iyH1(%esp)
        movapd %xmm2,nb211nf_izH1(%esp)
        movapd %xmm3,nb211nf_ixH2(%esp)
        movapd %xmm4,nb211nf_iyH2(%esp)
        movapd %xmm5,nb211nf_izH2(%esp)

        ## clear vctot 
        xorpd %xmm4,%xmm4
        movapd %xmm4,nb211nf_vctot(%esp)
        movapd %xmm4,nb211nf_Vvdwtot(%esp)

        movl  nb211nf_jindex(%ebp),%eax
        movl  (%eax,%esi,4),%ecx             ## jindex[n] 
        movl  4(%eax,%esi,4),%edx            ## jindex[n+1] 
        subl  %ecx,%edx              ## number of innerloop atoms 

        movl  nb211nf_pos(%ebp),%esi
        movl  nb211nf_jjnr(%ebp),%eax
        shll  $2,%ecx
        addl  %ecx,%eax
        movl  %eax,nb211nf_innerjjnr(%esp)       ## pointer to jjnr[nj0] 
        movl  %edx,%ecx
        subl  $2,%edx
        addl  nb211nf_ninner(%esp),%ecx
        movl  %ecx,nb211nf_ninner(%esp)
        addl  $0,%edx
        movl  %edx,nb211nf_innerk(%esp)      ## number of innerloop atoms 
        jge   _nb_kernel211nf_ia32_sse2.nb211nf_unroll_loop
        jmp   _nb_kernel211nf_ia32_sse2.nb211nf_checksingle
_nb_kernel211nf_ia32_sse2.nb211nf_unroll_loop: 
        ## twice unrolled innerloop here 
        movl  nb211nf_innerjjnr(%esp),%edx       ## pointer to jjnr[k] 
        movl  (%edx),%eax
        movl  4(%edx),%ebx

        addl $8,nb211nf_innerjjnr(%esp)                 ## advance pointer (unrolled 2) 

        movl nb211nf_charge(%ebp),%esi     ## base of charge[] 

        movlpd (%esi,%eax,8),%xmm3
        movhpd (%esi,%ebx,8),%xmm3
        movapd %xmm3,%xmm4
        mulpd  nb211nf_iqO(%esp),%xmm3
        mulpd  nb211nf_iqH(%esp),%xmm4

        movd  %eax,%mm0         ## use mmx registers as temp storage 
        movd  %ebx,%mm1

        movapd  %xmm3,nb211nf_qqO(%esp)
        movapd  %xmm4,nb211nf_qqH(%esp)

        movl nb211nf_type(%ebp),%esi
        movl (%esi,%eax,4),%eax
        movl (%esi,%ebx,4),%ebx
        movl nb211nf_vdwparam(%ebp),%esi
        shll %eax
        shll %ebx
        movl nb211nf_ntia(%esp),%edi
        addl %edi,%eax
        addl %edi,%ebx

        movlpd (%esi,%eax,8),%xmm6      ## c6a
        movlpd (%esi,%ebx,8),%xmm7      ## c6b
        movhpd 8(%esi,%eax,8),%xmm6     ## c6a c12a 
        movhpd 8(%esi,%ebx,8),%xmm7     ## c6b c12b 
        movapd %xmm6,%xmm4
        unpcklpd %xmm7,%xmm4
        unpckhpd %xmm7,%xmm6

        movd  %mm0,%eax
        movd  %mm1,%ebx
        movapd %xmm4,nb211nf_c6(%esp)
        movapd %xmm6,nb211nf_c12(%esp)

        movl nb211nf_pos(%ebp),%esi        ## base of pos[] 

        leal  (%eax,%eax,2),%eax     ## replace jnr with j3 
        leal  (%ebx,%ebx,2),%ebx

        ## move two coordinates to xmm0-xmm2 
        movlpd (%esi,%eax,8),%xmm0
        movlpd 8(%esi,%eax,8),%xmm1
        movlpd 16(%esi,%eax,8),%xmm2
        movhpd (%esi,%ebx,8),%xmm0
        movhpd 8(%esi,%ebx,8),%xmm1
        movhpd 16(%esi,%ebx,8),%xmm2

        ## move ixO-izO to xmm4-xmm6 
        movapd nb211nf_ixO(%esp),%xmm4
        movapd nb211nf_iyO(%esp),%xmm5
        movapd nb211nf_izO(%esp),%xmm6

        ## calc dr 
        subpd %xmm0,%xmm4
        subpd %xmm1,%xmm5
        subpd %xmm2,%xmm6

        ## square it 
        mulpd %xmm4,%xmm4
        mulpd %xmm5,%xmm5
        mulpd %xmm6,%xmm6
        addpd %xmm5,%xmm4
        addpd %xmm6,%xmm4
        movapd %xmm4,%xmm7
        ## rsqO in xmm7 

        ## move ixH1-izH1 to xmm4-xmm6 
        movapd nb211nf_ixH1(%esp),%xmm4
        movapd nb211nf_iyH1(%esp),%xmm5
        movapd nb211nf_izH1(%esp),%xmm6

        ## calc dr 
        subpd %xmm0,%xmm4
        subpd %xmm1,%xmm5
        subpd %xmm2,%xmm6

        ## square it 
        mulpd %xmm4,%xmm4
        mulpd %xmm5,%xmm5
        mulpd %xmm6,%xmm6
        addpd %xmm5,%xmm6
        addpd %xmm4,%xmm6
        ## rsqH1 in xmm6 

        ## move ixH2-izH2 to xmm3-xmm5  
        movapd nb211nf_ixH2(%esp),%xmm3
        movapd nb211nf_iyH2(%esp),%xmm4
        movapd nb211nf_izH2(%esp),%xmm5

        ## calc dr 
        subpd %xmm0,%xmm3
        subpd %xmm1,%xmm4
        subpd %xmm2,%xmm5

        ## square it 
        mulpd %xmm3,%xmm3
        mulpd %xmm4,%xmm4
        mulpd %xmm5,%xmm5
        addpd %xmm4,%xmm5
        addpd %xmm3,%xmm5
        ## rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 

        movapd %xmm5,%xmm0
        movapd %xmm6,%xmm1
        movapd %xmm7,%xmm2

        mulpd  nb211nf_krf(%esp),%xmm0
        mulpd  nb211nf_krf(%esp),%xmm1
        mulpd  nb211nf_krf(%esp),%xmm2

        movapd %xmm0,nb211nf_krsqH2(%esp)
        movapd %xmm1,nb211nf_krsqH1(%esp)
        movapd %xmm2,nb211nf_krsqO(%esp)

        ## start with rsqO - put seed in xmm2 
        cvtpd2ps %xmm7,%xmm2
        rsqrtps %xmm2,%xmm2
        cvtps2pd %xmm2,%xmm2

        movapd  %xmm2,%xmm3
        mulpd   %xmm2,%xmm2