nb_kernel303_x86_64_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 1,919 行 · 第 1/4 页

        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 
        movaps %xmm0,%xmm6
        movss %xmm2,%xmm7
        movlhps %xmm1,%xmm6
        shufps $8,%xmm6,%xmm6 ## 00001000       

        ## accumulate H2i forces in xmm0, xmm1, xmm2 
        movaps nb303_fixH2(%rsp),%xmm0
        movaps nb303_fiyH2(%rsp),%xmm1
        movaps nb303_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

        ## accumulate M i forces in xmm0, xmm1, xmm2 
        movaps nb303_fixM(%rsp),%xmm0
        movaps nb303_fiyM(%rsp),%xmm1
        movaps nb303_fizM(%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  36(%rdi,%rcx,4),%xmm3
        movss  40(%rdi,%rcx,4),%xmm4
        movss  44(%rdi,%rcx,4),%xmm5
        subss  %xmm0,%xmm3
        subss  %xmm1,%xmm4
        subss  %xmm2,%xmm5
        movss  %xmm3,36(%rdi,%rcx,4)
        movss  %xmm4,40(%rdi,%rcx,4)
        movss  %xmm5,44(%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 nb303_n(%rsp),%esi
        ## get group index for i particle 
        movq  nb303_gid(%rbp),%rdx              ## base of gid[]
        movl  (%rdx,%rsi,4),%edx                ## ggid=gid[n]

        ## accumulate total potential energy and update it 
        movaps nb303_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  nb303_Vc(%rbp),%rax
        addss (%rax,%rdx,4),%xmm7
        ## move back to mem 
        movss %xmm7,(%rax,%rdx,4)

        ## finish if last 
        movl nb303_nn1(%rsp),%ecx
        ## esi already loaded with n
        incl %esi
        subl %esi,%ecx
        jz _nb_kernel303_x86_64_sse.nb303_outerend

        ## not last, iterate outer loop once more!  
        movl %esi,nb303_n(%rsp)
        jmp _nb_kernel303_x86_64_sse.nb303_outer
_nb_kernel303_x86_64_sse.nb303_outerend: 
        ## check if more outer neighborlists remain
        movl  nb303_nri(%rsp),%ecx
        ## esi already loaded with n above
        subl  %esi,%ecx
        jz _nb_kernel303_x86_64_sse.nb303_end
        ## non-zero, do one more workunit
        jmp   _nb_kernel303_x86_64_sse.nb303_threadloop
_nb_kernel303_x86_64_sse.nb303_end: 
        movl nb303_nouter(%rsp),%eax
        movl nb303_ninner(%rsp),%ebx
        movq nb303_outeriter(%rbp),%rcx
        movq nb303_inneriter(%rbp),%rdx
        movl %eax,(%rcx)
        movl %ebx,(%rdx)

        addq $824,%rsp
        emms


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

        pop %rbx
        pop    %rbp
        ret




.globl nb_kernel303nf_x86_64_sse
.globl _nb_kernel303nf_x86_64_sse
nb_kernel303nf_x86_64_sse:      
_nb_kernel303nf_x86_64_sse:     
##      Room for return address and rbp (16 bytes)
.set nb303nf_fshift, 16
.set nb303nf_gid, 24
.set nb303nf_pos, 32
.set nb303nf_faction, 40
.set nb303nf_charge, 48
.set nb303nf_p_facel, 56
.set nb303nf_argkrf, 64
.set nb303nf_argcrf, 72
.set nb303nf_Vc, 80
.set nb303nf_type, 88
.set nb303nf_p_ntype, 96
.set nb303nf_vdwparam, 104
.set nb303nf_Vvdw, 112
.set nb303nf_p_tabscale, 120
.set nb303nf_VFtab, 128
.set nb303nf_invsqrta, 136
.set nb303nf_dvda, 144
.set nb303nf_p_gbtabscale, 152
.set nb303nf_GBtab, 160
.set nb303nf_p_nthreads, 168
.set nb303nf_count, 176
.set nb303nf_mtx, 184
.set nb303nf_outeriter, 192
.set nb303nf_inneriter, 200
.set nb303nf_work, 208
        ## stack offsets for local variables  
        ## bottom of stack is cache-aligned for sse use 
.set nb303nf_ixH1, 0
.set nb303nf_iyH1, 16
.set nb303nf_izH1, 32
.set nb303nf_ixH2, 48
.set nb303nf_iyH2, 64
.set nb303nf_izH2, 80
.set nb303nf_ixM, 96
.set nb303nf_iyM, 112
.set nb303nf_izM, 128
.set nb303nf_iqH, 144
.set nb303nf_iqM, 160
.set nb303nf_qqH, 176
.set nb303nf_qqM, 192
.set nb303nf_rinvH1, 208
.set nb303nf_rinvH2, 224
.set nb303nf_rinvM, 240
.set nb303nf_rH1, 256
.set nb303nf_rH2, 272
.set nb303nf_rM, 288
.set nb303nf_tsc, 304
.set nb303nf_vctot, 320
.set nb303nf_half, 336
.set nb303nf_three, 352
.set nb303nf_is3, 368
.set nb303nf_ii3, 372
.set nb303nf_nri, 376
.set nb303nf_iinr, 384
.set nb303nf_jindex, 392
.set nb303nf_jjnr, 400
.set nb303nf_shift, 408
.set nb303nf_shiftvec, 416
.set nb303nf_facel, 424
.set nb303nf_innerjjnr, 432
.set nb303nf_innerk, 440
.set nb303nf_n, 444
.set nb303nf_nn1, 448
.set nb303nf_nouter, 452
.set nb303nf_ninner, 456
        push %rbp
        movq %rsp,%rbp
        push %rbx

        emms

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

        subq $472,%rsp          ## local variable stack space (n*16+8)

        ## zero 32-bit iteration counters
        movl $0,%eax
        movl %eax,nb303nf_nouter(%rsp)
        movl %eax,nb303nf_ninner(%rsp)

        movl (%rdi),%edi
        movl %edi,nb303nf_nri(%rsp)
        movq %rsi,nb303nf_iinr(%rsp)
        movq %rdx,nb303nf_jindex(%rsp)
        movq %rcx,nb303nf_jjnr(%rsp)
        movq %r8,nb303nf_shift(%rsp)
        movq %r9,nb303nf_shiftvec(%rsp)
        movq nb303nf_p_facel(%rbp),%rsi
        movss (%rsi),%xmm0
        movss %xmm0,nb303nf_facel(%rsp)

        movq nb303nf_p_tabscale(%rbp),%rax
        movss (%rax),%xmm3
        shufps $0,%xmm3,%xmm3
        movaps %xmm3,nb303nf_tsc(%rsp)

        ## create constant floating-point factors on stack
        movl $0x3f000000,%eax   ## half in IEEE (hex)
        movl %eax,nb303nf_half(%rsp)
        movss nb303nf_half(%rsp),%xmm1
        shufps $0,%xmm1,%xmm1  ## splat to all elements
        movaps %xmm1,%xmm2
        addps  %xmm2,%xmm2      ## one
        movaps %xmm2,%xmm3
        addps  %xmm2,%xmm2      ## two
        addps  %xmm2,%xmm3      ## three
        movaps %xmm1,nb303nf_half(%rsp)
        movaps %xmm3,nb303nf_three(%rsp)

        ## assume we have at least one i particle - start directly 
        movq  nb303nf_iinr(%rsp),%rcx           ## rcx = pointer into iinr[]    
        movl  (%rcx),%ebx               ## ebx =ii 

        movq  nb303nf_charge(%rbp),%rdx
        movss 4(%rdx,%rbx,4),%xmm3
        movss 12(%rdx,%rbx,4),%xmm4
        movq nb303nf_p_facel(%rbp),%rsi
        movss (%rsi),%xmm0
        movss nb303nf_facel(%rsp),%xmm5
        mulss  %xmm5,%xmm3
        mulss  %xmm5,%xmm4

        shufps $0,%xmm3,%xmm3
        shufps $0,%xmm4,%xmm4
        movaps %xmm3,nb303nf_iqH(%rsp)
        movaps %xmm4,nb303nf_iqM(%rsp)

_nb_kernel303nf_x86_64_sse.nb303nf_threadloop: 
        movq  nb303nf_count(%rbp),%rsi            ## pointer to sync counter
        movl  (%rsi),%eax
_nb_kernel303nf_x86_64_sse.nb303nf_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_kernel303nf_x86_64_sse.nb303nf_spinlock

        ## if(nn1>nri) nn1=nri
        movl nb303nf_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,nb303nf_n(%rsp)
        movl %ebx,nb303nf_nn1(%rsp)
        subl %eax,%ebx                          ## calc number of outer lists
        movl %eax,%esi                          ## copy n to esi
        jg  _nb_kernel303nf_x86_64_sse.nb303nf_outerstart
        jmp _nb_kernel303nf_x86_64_sse.nb303nf_end

_nb_kernel303nf_x86_64_sse.nb303nf_outerstart: 
        ## ebx contains number of outer iterations
        addl nb303nf_nouter(%rsp),%ebx
        movl %ebx,nb303nf_nouter(%rsp)

_nb_kernel303nf_x86_64_sse.nb303nf_outer: 
        movq  nb303nf_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,nb303nf_is3(%rsp)            ## store is3 

        movq  nb303nf_shiftvec(%rsp),%rax     ## rax = base of shiftvec[] 

        movss (%rax,%rbx,4),%xmm0
        movss 4(%rax,%rbx,4),%xmm1
        movss 8(%rax,%rbx,4),%xmm2

        movq  nb303nf_iinr(%rsp),%rcx           ## rcx = pointer into iinr[]    
        movl  (%rcx,%rsi,4),%ebx                ## ebx =ii 

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

        lea  (%rbx,%rbx,2),%rbx        ## rbx = 3*ii=ii3 
        movq  nb303nf_pos(%rbp),%rax    ## rax = base of pos[]  
        movl  %ebx,nb303nf_ii3(%rsp)

        addss 12(%rax,%rbx,4),%xmm3
        addss 16(%rax,%rbx,4),%xmm4
        addss 20(%rax,%rbx,4),%xmm5
        shufps $0,%xmm3,%xmm3
        shufps $0,%xmm4,%xmm4
        shufps $0,%xmm5,%xmm5
        movaps %xmm3,nb303nf_ixH1(%rsp)
        movaps %xmm4,nb303nf_iyH1(%rsp)
        movaps %xmm5,nb303nf_izH1(%rsp)

        movss %xmm0,%xmm3
        movss %xmm1,%xmm4
        movss %xmm2,%xmm5
        addss 24(%rax,%rbx,4),%xmm0
        addss 28(%rax,%rbx,4),%xmm1
        addss 32(%rax,%rbx,4),%xmm2
        addss 36(%rax,%rbx,4),%xmm3
        addss 40(%rax,%rbx,4),%xmm4
        addss 44(%rax,%rbx,4),%xmm5

        shufps $0,%xmm0,%xmm0
        shufps $0,%xmm1,%xmm1
        shufps $0,%xmm2,%xmm2
        shufps $0,%xmm3,%xmm3
        shufps $0,%xmm4,%xmm4
        shufps $0,%xmm5,%xmm5
        movaps %xmm0,nb303nf_ixH2(%rsp)
        movaps %xmm1,nb303nf_iyH2(%rsp)
        movaps %xmm2,nb303nf_izH2(%rsp)
        movaps %xmm3,nb303nf_ixM(%rsp)
        movaps %xmm4,nb303nf_iyM(%rsp)
        movaps %xmm5,nb303nf_izM(%rsp)

        ## clear vctot 
        xorps %xmm4,%xmm4
        movaps %xmm4,nb303nf_vctot(%rsp)

        movq  nb303nf_jindex(%rsp),%rax
        movl  (%rax,%rsi,4),%ecx                ## jindex[n] 
        movl  4(%rax,%rsi,4),%edx               ## jindex[n+1] 
        subl  %ecx,%edx                 ## number of innerloop atoms 

        movq  nb303nf_pos(%rbp),%rsi
        movq  nb303nf_faction(%rbp),%rdi
        movq  nb303nf_jjnr(%rsp),%rax
        shll  $2,%ecx
        addq  %rcx,%rax
        movq  %rax,nb303nf_innerjjnr(%rsp)      ## pointer to jjnr[nj0] 
        movl  %edx,%ecx
        subl  $4,%edx
        addl  nb303nf_ninner(%rsp),%ecx
        movl  %ecx,nb303nf_ninner(%rsp)
        addl  $0,%edx
        movl  %edx,nb303nf_innerk(%rsp)         ## number of innerloop atoms 
        jge   _nb_kernel303nf_x86_64_sse.nb303nf_unroll_loop
        jmp   _nb_kernel303nf_x86_64_sse.nb303nf_odd_inner
_nb_kernel303nf_x86_64_sse.nb303nf_unroll_loop: 
        ## quad-unroll innerloop here 
        movq  nb303nf_innerjjnr(%rsp),%rdx      ## pointer to jjnr[k] 
        movl  (%rdx),%eax
        movl  4(%rdx),%ebx
        movl  8(%rdx),%ecx
        movl  12(%rdx),%edx             ## eax-edx=jnr1-4 

        addq $16,nb303nf_innerjjnr(%rsp)             ## advance pointer (unrolled 4) 

        movq nb303nf_charge(%rbp),%rsi  ## base of charge[] 

        movss (%rsi,%rax,4),%xmm3
        movss (%rsi,%rcx,4),%xmm4
        movss (%rsi,%rbx,4),%xmm6
        movss (%rsi,%rdx,4),%xmm7

        shufps $0,%xmm6,%xmm3
        shufps $0,%xmm7,%xmm4
        shufps $136,%xmm4,%xmm3 ## 10001000 ;# all charges in xmm3  
        movaps %xmm3,%xmm4              ## and in xmm4 
        mulps  nb303nf_iqH(%rsp),%xmm3
        mulps  nb303nf_iqM(%rsp),%xmm4

        movd  %eax,%mm0         ## use mmx registers as temp storage 
        movd  %ebx,%mm1
        movd  %ecx,%mm2
        movd  %edx,%mm3

        movaps  %xmm3,nb303nf_qqH(%rsp)
        movaps  %xmm4,nb303nf_qqM(%rsp)

        movq nb303nf_pos(%rbp),%rsi     ## base of pos[] 

        lea  (%rax,%rax,2),%rax        ## replace jnr with j3 
        lea  (%rbx,%rbx,2),%rbx
        lea  (%rcx,%rcx,2),%rcx        ## replace jnr with j3 
        lea  (%rdx,%rdx,2),%rdx

        ## move four coordinates to xmm0-xmm2   
        movlps (%rsi,%rax,4),%xmm4
        movlps (%rsi,%rcx,4),%xmm5
        movss 8(%rsi,%rax,4),%xmm2
        movss 8(%rsi,%rcx,4),%xmm6

        movhps (%rsi,%rbx,4),%xmm4
        movhps (%rsi,%rdx,4),%xmm5

        movss 8(%rsi,%rbx,4),%xmm0
        movss 8(%rsi,%rdx,4),%xmm1

        shufps $0,%xmm0,%xmm2