nb_kernel313_x86_64_sse2.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,129 行 · 第 1/5 页

        movlpd %xmm7,nb313_fizM(%rsp)

        movq nb313_faction(%rbp),%rdi
        ## update j forces 
        addsd  nb313_fjx(%rsp),%xmm0
        addsd  nb313_fjy(%rsp),%xmm1
        addsd  nb313_fjz(%rsp),%xmm2

        ## the fj's - start by accumulating forces from memory 
        movlpd (%rdi,%rax,8),%xmm3
        movlpd 8(%rdi,%rax,8),%xmm4
        movlpd 16(%rdi,%rax,8),%xmm5
        addsd %xmm0,%xmm3
        addsd %xmm1,%xmm4
        addsd %xmm2,%xmm5
        movlpd %xmm3,(%rdi,%rax,8)
        movlpd %xmm4,8(%rdi,%rax,8)
        movlpd %xmm5,16(%rdi,%rax,8)

_nb_kernel313_x86_64_sse2.nb313_updateouterdata: 
        movl  nb313_ii3(%rsp),%ecx
        movq  nb313_faction(%rbp),%rdi
        movq  nb313_fshift(%rbp),%rsi
        movl  nb313_is3(%rsp),%edx

        ## accumulate  Oi forces in xmm0, xmm1, xmm2 
        movapd nb313_fixO(%rsp),%xmm0
        movapd nb313_fiyO(%rsp),%xmm1
        movapd nb313_fizO(%rsp),%xmm2

        movhlps %xmm0,%xmm3
        movhlps %xmm1,%xmm4
        movhlps %xmm2,%xmm5
        addsd  %xmm3,%xmm0
        addsd  %xmm4,%xmm1
        addsd  %xmm5,%xmm2 ## sum is in low xmm0-xmm2 

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

        ## increment i force 
        movsd  (%rdi,%rcx,8),%xmm3
        movsd  8(%rdi,%rcx,8),%xmm4
        movsd  16(%rdi,%rcx,8),%xmm5
        subsd  %xmm0,%xmm3
        subsd  %xmm1,%xmm4
        subsd  %xmm2,%xmm5
        movsd  %xmm3,(%rdi,%rcx,8)
        movsd  %xmm4,8(%rdi,%rcx,8)
        movsd  %xmm5,16(%rdi,%rcx,8)

        ## accumulate force in xmm6/xmm7 for fshift 
        movapd %xmm0,%xmm6
        movsd %xmm2,%xmm7
        unpcklpd %xmm1,%xmm6

        ## accumulate H1i forces in xmm0, xmm1, xmm2 
        movapd nb313_fixH1(%rsp),%xmm0
        movapd nb313_fiyH1(%rsp),%xmm1
        movapd nb313_fizH1(%rsp),%xmm2

        movhlps %xmm0,%xmm3
        movhlps %xmm1,%xmm4
        movhlps %xmm2,%xmm5
        addsd  %xmm3,%xmm0
        addsd  %xmm4,%xmm1
        addsd  %xmm5,%xmm2 ## sum is in low xmm0-xmm2 

        ## increment i force 
        movsd  24(%rdi,%rcx,8),%xmm3
        movsd  32(%rdi,%rcx,8),%xmm4
        movsd  40(%rdi,%rcx,8),%xmm5
        subsd  %xmm0,%xmm3
        subsd  %xmm1,%xmm4
        subsd  %xmm2,%xmm5
        movsd  %xmm3,24(%rdi,%rcx,8)
        movsd  %xmm4,32(%rdi,%rcx,8)
        movsd  %xmm5,40(%rdi,%rcx,8)

        ## accumulate force in xmm6/xmm7 for fshift 
        addsd %xmm2,%xmm7
        unpcklpd %xmm1,%xmm0
        addpd %xmm0,%xmm6

        ## accumulate H2i forces in xmm0, xmm1, xmm2 
        movapd nb313_fixH2(%rsp),%xmm0
        movapd nb313_fiyH2(%rsp),%xmm1
        movapd nb313_fizH2(%rsp),%xmm2

        movhlps %xmm0,%xmm3
        movhlps %xmm1,%xmm4
        movhlps %xmm2,%xmm5
        addsd  %xmm3,%xmm0
        addsd  %xmm4,%xmm1
        addsd  %xmm5,%xmm2 ## sum is in low xmm0-xmm2 

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

        ## increment i force 
        movsd  48(%rdi,%rcx,8),%xmm3
        movsd  56(%rdi,%rcx,8),%xmm4
        movsd  64(%rdi,%rcx,8),%xmm5
        subsd  %xmm0,%xmm3
        subsd  %xmm1,%xmm4
        subsd  %xmm2,%xmm5
        movsd  %xmm3,48(%rdi,%rcx,8)
        movsd  %xmm4,56(%rdi,%rcx,8)
        movsd  %xmm5,64(%rdi,%rcx,8)

        ## accumulate force in xmm6/xmm7 for fshift 
        addsd %xmm2,%xmm7
        unpcklpd %xmm1,%xmm0
        addpd %xmm0,%xmm6

        ## accumulate Mi forces in xmm0, xmm1, xmm2 
        movapd nb313_fixM(%rsp),%xmm0
        movapd nb313_fiyM(%rsp),%xmm1
        movapd nb313_fizM(%rsp),%xmm2

        movhlps %xmm0,%xmm3
        movhlps %xmm1,%xmm4
        movhlps %xmm2,%xmm5
        addsd  %xmm3,%xmm0
        addsd  %xmm4,%xmm1
        addsd  %xmm5,%xmm2 ## sum is in low xmm0-xmm2 

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

        ## increment i force 
        movsd  72(%rdi,%rcx,8),%xmm3
        movsd  80(%rdi,%rcx,8),%xmm4
        movsd  88(%rdi,%rcx,8),%xmm5
        subsd  %xmm0,%xmm3
        subsd  %xmm1,%xmm4
        subsd  %xmm2,%xmm5
        movsd  %xmm3,72(%rdi,%rcx,8)
        movsd  %xmm4,80(%rdi,%rcx,8)
        movsd  %xmm5,88(%rdi,%rcx,8)

        ## accumulate force in xmm6/xmm7 for fshift 
        addsd %xmm2,%xmm7
        unpcklpd %xmm1,%xmm0
        addpd %xmm0,%xmm6

        ## increment fshift force 
        movlpd (%rsi,%rdx,8),%xmm3
        movhpd 8(%rsi,%rdx,8),%xmm3
        movsd  16(%rsi,%rdx,8),%xmm4
        subpd  %xmm6,%xmm3
        subsd  %xmm7,%xmm4
        movlpd %xmm3,(%rsi,%rdx,8)
        movhpd %xmm3,8(%rsi,%rdx,8)
        movsd  %xmm4,16(%rsi,%rdx,8)

        ## get n from stack
        movl nb313_n(%rsp),%esi
        ## get group index for i particle 
        movq  nb313_gid(%rbp),%rdx              ## base of gid[]
        movl  (%rdx,%rsi,4),%edx                ## ggid=gid[n]

        ## accumulate total potential energy and update it 
        movapd nb313_vctot(%rsp),%xmm7
        ## accumulate 
        movhlps %xmm7,%xmm6
        addsd  %xmm6,%xmm7      ## low xmm7 has the sum now 

        ## add earlier value from mem 
        movq  nb313_Vc(%rbp),%rax
        addsd (%rax,%rdx,8),%xmm7
        ## move back to mem 
        movsd %xmm7,(%rax,%rdx,8)

        ## accumulate total lj energy and update it 
        movapd nb313_Vvdwtot(%rsp),%xmm7
        ## accumulate 
        movhlps %xmm7,%xmm6
        addsd  %xmm6,%xmm7      ## low xmm7 has the sum now 

        ## add earlier value from mem 
        movq  nb313_Vvdw(%rbp),%rax
        addsd (%rax,%rdx,8),%xmm7
        ## move back to mem 
        movsd %xmm7,(%rax,%rdx,8)

        ## finish if last 
        movl nb313_nn1(%rsp),%ecx
        ## esi already loaded with n
        incl %esi
        subl %esi,%ecx
        jz _nb_kernel313_x86_64_sse2.nb313_outerend

        ## not last, iterate outer loop once more!  
        movl %esi,nb313_n(%rsp)
        jmp _nb_kernel313_x86_64_sse2.nb313_outer
_nb_kernel313_x86_64_sse2.nb313_outerend: 
        ## check if more outer neighborlists remain
        movl  nb313_nri(%rsp),%ecx
        ## esi already loaded with n above
        subl  %esi,%ecx
        jz _nb_kernel313_x86_64_sse2.nb313_end
        ## non-zero, do one more workunit
        jmp   _nb_kernel313_x86_64_sse2.nb313_threadloop
_nb_kernel313_x86_64_sse2.nb313_end: 
        movl nb313_nouter(%rsp),%eax
        movl nb313_ninner(%rsp),%ebx
        movq nb313_outeriter(%rbp),%rcx
        movq nb313_inneriter(%rbp),%rdx
        movl %eax,(%rcx)
        movl %ebx,(%rdx)

        addq $1080,%rsp
        emms


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

        pop %rbx
        pop    %rbp
        ret






.globl nb_kernel313nf_x86_64_sse2
.globl _nb_kernel313nf_x86_64_sse2
nb_kernel313nf_x86_64_sse2:     
_nb_kernel313nf_x86_64_sse2:    
##      Room for return address and rbp (16 bytes)
.set nb313nf_fshift, 16
.set nb313nf_gid, 24
.set nb313nf_pos, 32
.set nb313nf_faction, 40
.set nb313nf_charge, 48
.set nb313nf_p_facel, 56
.set nb313nf_argkrf, 64
.set nb313nf_argcrf, 72
.set nb313nf_Vc, 80
.set nb313nf_type, 88
.set nb313nf_p_ntype, 96
.set nb313nf_vdwparam, 104
.set nb313nf_Vvdw, 112
.set nb313nf_p_tabscale, 120
.set nb313nf_VFtab, 128
.set nb313nf_invsqrta, 136
.set nb313nf_dvda, 144
.set nb313nf_p_gbtabscale, 152
.set nb313nf_GBtab, 160
.set nb313nf_p_nthreads, 168
.set nb313nf_count, 176
.set nb313nf_mtx, 184
.set nb313nf_outeriter, 192
.set nb313nf_inneriter, 200
.set nb313nf_work, 208
        ## stack offsets for local variables 
        ## bottom of stack is cache-aligned for sse2 use 
.set nb313nf_ixO, 0
.set nb313nf_iyO, 16
.set nb313nf_izO, 32
.set nb313nf_ixH1, 48
.set nb313nf_iyH1, 64
.set nb313nf_izH1, 80
.set nb313nf_ixH2, 96
.set nb313nf_iyH2, 112
.set nb313nf_izH2, 128
.set nb313nf_ixM, 144
.set nb313nf_iyM, 160
.set nb313nf_izM, 176
.set nb313nf_iqM, 192
.set nb313nf_iqH, 208
.set nb313nf_qqM, 224
.set nb313nf_qqH, 240
.set nb313nf_rinvsqO, 256
.set nb313nf_rinvH1, 272
.set nb313nf_rinvH2, 288
.set nb313nf_rinvM, 304
.set nb313nf_rO, 320
.set nb313nf_rH1, 336
.set nb313nf_rH2, 352
.set nb313nf_rM, 368
.set nb313nf_tsc, 384
.set nb313nf_two, 400
.set nb313nf_c6, 416
.set nb313nf_c12, 432
.set nb313nf_vctot, 448
.set nb313nf_Vvdwtot, 464
.set nb313nf_half, 480
.set nb313nf_three, 496
.set nb313nf_is3, 512
.set nb313nf_ii3, 516
.set nb313nf_nri, 520
.set nb313nf_iinr, 528
.set nb313nf_jindex, 536
.set nb313nf_jjnr, 544
.set nb313nf_shift, 552
.set nb313nf_shiftvec, 560
.set nb313nf_facel, 568
.set nb313nf_innerjjnr, 576
.set nb313nf_ntia, 584
.set nb313nf_innerk, 588
.set nb313nf_n, 592
.set nb313nf_nn1, 596
.set nb313nf_nouter, 600
.set nb313nf_ninner, 604
        push %rbp
        movq %rsp,%rbp
        push %rbx
        emms

        push %r12
        push %r13
        push %r14
        push %r15

        subq $616,%rsp          ## local variable stack space (n*16+8)

        ## zero 32-bit iteration counters
        movl $0,%eax
        movl %eax,nb313nf_nouter(%rsp)
        movl %eax,nb313nf_ninner(%rsp)

        movl (%rdi),%edi
        movl %edi,nb313nf_nri(%rsp)
        movq %rsi,nb313nf_iinr(%rsp)
        movq %rdx,nb313nf_jindex(%rsp)
        movq %rcx,nb313nf_jjnr(%rsp)
        movq %r8,nb313nf_shift(%rsp)
        movq %r9,nb313nf_shiftvec(%rsp)
        movq nb313nf_p_facel(%rbp),%rsi
        movsd (%rsi),%xmm0
        movsd %xmm0,nb313nf_facel(%rsp)

        movq nb313nf_p_tabscale(%rbp),%rax
        movsd (%rax),%xmm3
        shufpd $0,%xmm3,%xmm3
        movapd %xmm3,nb313nf_tsc(%rsp)

        ## create constant floating-point factors on stack
        movl $0x00000000,%eax   ## lower half of double half IEEE (hex)
        movl $0x3fe00000,%ebx
        movl %eax,nb313nf_half(%rsp)
        movl %ebx,nb313nf_half+4(%rsp)
        movsd nb313nf_half(%rsp),%xmm1
        shufpd $0,%xmm1,%xmm1  ## splat to all elements
        movapd %xmm1,%xmm3
        addpd  %xmm3,%xmm3      ## one
        movapd %xmm3,%xmm2
        addpd  %xmm2,%xmm2      ## two
        addpd  %xmm2,%xmm3      ## three
        movapd %xmm1,nb313nf_half(%rsp)
        movapd %xmm2,nb313nf_two(%rsp)
        movapd %xmm3,nb313nf_three(%rsp)

        ## assume we have at least one i particle - start directly 
        movq  nb313nf_iinr(%rsp),%rcx         ## rcx = pointer into iinr[]      
        movl  (%rcx),%ebx           ## ebx =ii 

        movq  nb313nf_charge(%rbp),%rdx
        movsd 8(%rdx,%rbx,8),%xmm3
        movsd 24(%rdx,%rbx,8),%xmm4
        movq nb313nf_p_facel(%rbp),%rsi
        movsd (%rsi),%xmm0
        movsd nb313nf_facel(%rsp),%xmm5
        mulsd  %xmm5,%xmm3
        mulsd  %xmm5,%xmm4

        shufpd $0,%xmm3,%xmm3
        shufpd $0,%xmm4,%xmm4
        movapd %xmm3,nb313nf_iqH(%rsp)
        movapd %xmm4,nb313nf_iqM(%rsp)

        movq  nb313nf_type(%rbp),%rdx
        movl  (%rdx,%rbx,4),%ecx
        shll  %ecx
        movq nb313nf_p_ntype(%rbp),%rdi
        imull (%rdi),%ecx     ## rcx = ntia = 2*ntype*type[ii0] 
        movl  %ecx,nb313nf_ntia(%rsp)
_nb_kernel313nf_x86_64_sse2.nb313nf_threadloop: 
        movq  nb313nf_count(%rbp),%rsi            ## pointer to sync counter
        movl  (%rsi),%eax
_nb_kernel313nf_x86_64_sse2.nb313nf_spinlock: 
        movl  %eax,%ebx                         ## ebx=*count=nn0
        addl  $1,%ebx                          ## ebx=nn1=nn0+10
        lock 
        cmpxchgl %ebx,(%rsi)                    ## write nn1 to *counter,
                                                ## if it hasnt changed.
                                                ## or reread *counter to eax.
        pause                                   ## -> better p4 performance
        jnz _nb_kernel313nf_x86_64_sse2.nb313nf_spinlock

        ## if(nn1>nri) nn1=nri
        movl nb313nf_nri(%rsp),%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,nb313nf_n(%rsp)
        movl %ebx,nb313nf_nn1(%rsp)
        subl %eax,%ebx                          ## calc number of outer lists
        movl %eax,%esi                          ## copy n to esi
        jg  _nb_kernel313nf_x86_64_sse2.nb313nf_outerstart
        jmp _nb_kernel313nf_x86_64_sse2.nb313nf_end

_nb_kernel313nf_x86_64_sse2.nb313nf_outerstart: 
        ## ebx contains number of outer iterations
        addl nb313nf_nouter(%rsp),%ebx
        movl %ebx,nb313nf_nouter(%rsp)

_nb_kernel313nf_x86_64_sse2.nb313nf_outer: 
        movq  nb313nf_shift(%rsp),%rax        ## rax = pointer into shift[] 
        movl  (%rax,%rsi,4),%ebx        ## rbx=shift[n] 

        lea  (%rbx,%rbx,2),%rbx    ## rbx=3*is 
        movl  %ebx,nb313nf_is3(%rsp)            ## store is3 

        movq  nb313nf_shiftvec(%rsp),%rax     ## rax = base of shiftvec[]