nb_kernel211_x86_64_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 1,749 行 · 第 1/4 页

        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 nb211_fixH1(%rsp),%xmm0
        movaps nb211_fiyH1(%rsp),%xmm1
        movaps nb211_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 nb211_fixH2(%rsp),%xmm0
        movaps nb211_fiyH2(%rsp),%xmm1
        movaps nb211_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 nb211_n(%rsp),%esi
        ## get group index for i particle 
        movq  nb211_gid(%rbp),%rdx              ## base of gid[]
        movl  (%rdx,%rsi,4),%edx                ## ggid=gid[n]

        ## accumulate total potential energy and update it 
        movaps nb211_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  nb211_Vc(%rbp),%rax
        addss (%rax,%rdx,4),%xmm7
        ## move back to mem 
        movss %xmm7,(%rax,%rdx,4)

        ## accumulate total lj energy and update it 
        movaps nb211_Vvdwtot(%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  nb211_Vvdw(%rbp),%rax
        addss (%rax,%rdx,4),%xmm7
        ## move back to mem 
        movss %xmm7,(%rax,%rdx,4)

        ## finish if last 
        movl nb211_nn1(%rsp),%ecx
        ## esi already loaded with n
        incl %esi
        subl %esi,%ecx
        jz _nb_kernel211_x86_64_sse.nb211_outerend

        ## not last, iterate outer loop once more!  
        movl %esi,nb211_n(%rsp)
        jmp _nb_kernel211_x86_64_sse.nb211_outer
_nb_kernel211_x86_64_sse.nb211_outerend: 
        ## check if more outer neighborlists remain
        movl  nb211_nri(%rsp),%ecx
        ## esi already loaded with n above
        subl  %esi,%ecx
        jz _nb_kernel211_x86_64_sse.nb211_end
        ## non-zero, do one more workunit
        jmp   _nb_kernel211_x86_64_sse.nb211_threadloop
_nb_kernel211_x86_64_sse.nb211_end: 
        movl nb211_nouter(%rsp),%eax
        movl nb211_ninner(%rsp),%ebx
        movq nb211_outeriter(%rbp),%rcx
        movq nb211_inneriter(%rbp),%rdx
        movl %eax,(%rcx)
        movl %ebx,(%rdx)

        addq $872,%rsp
        emms


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

        pop %rbx
        pop    %rbp
        ret




.globl nb_kernel211nf_x86_64_sse
.globl _nb_kernel211nf_x86_64_sse
nb_kernel211nf_x86_64_sse:      
_nb_kernel211nf_x86_64_sse:     
##      Room for return address and rbp (16 bytes)
.set nb211nf_fshift, 16
.set nb211nf_gid, 24
.set nb211nf_pos, 32
.set nb211nf_faction, 40
.set nb211nf_charge, 48
.set nb211nf_p_facel, 56
.set nb211nf_argkrf, 64
.set nb211nf_argcrf, 72
.set nb211nf_Vc, 80
.set nb211nf_type, 88
.set nb211nf_p_ntype, 96
.set nb211nf_vdwparam, 104
.set nb211nf_Vvdw, 112
.set nb211nf_p_tabscale, 120
.set nb211nf_VFtab, 128
.set nb211nf_invsqrta, 136
.set nb211nf_dvda, 144
.set nb211nf_p_gbtabscale, 152
.set nb211nf_GBtab, 160
.set nb211nf_p_nthreads, 168
.set nb211nf_count, 176
.set nb211nf_mtx, 184
.set nb211nf_outeriter, 192
.set nb211nf_inneriter, 200
.set nb211nf_work, 208
        ## 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_nri, 384
.set nb211nf_iinr, 392
.set nb211nf_jindex, 400
.set nb211nf_jjnr, 408
.set nb211nf_shift, 416
.set nb211nf_shiftvec, 424
.set nb211nf_facel, 432
.set nb211nf_innerjjnr, 440
.set nb211nf_ntia, 448
.set nb211nf_is3, 452
.set nb211nf_ii3, 456
.set nb211nf_innerk, 460
.set nb211nf_n, 464
.set nb211nf_nn1, 468
.set nb211nf_nouter, 472
.set nb211nf_ninner, 476

        push %rbp
        movq %rsp,%rbp
        push %rbx

        emms

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

        subq $488,%rsp          ## local variable stack space (n*16+8)

        ## zero 32-bit iteration counters
        movl $0,%eax
        movl %eax,nb211nf_nouter(%rsp)
        movl %eax,nb211nf_ninner(%rsp)

        movl (%rdi),%edi
        movl %edi,nb211nf_nri(%rsp)
        movq %rsi,nb211nf_iinr(%rsp)
        movq %rdx,nb211nf_jindex(%rsp)
        movq %rcx,nb211nf_jjnr(%rsp)
        movq %r8,nb211nf_shift(%rsp)
        movq %r9,nb211nf_shiftvec(%rsp)
        movq nb211nf_p_facel(%rbp),%rsi
        movss (%rsi),%xmm0
        movss %xmm0,nb211nf_facel(%rsp)


        movq nb211nf_argkrf(%rbp),%rsi
        movq nb211nf_argcrf(%rbp),%rdi
        movss (%rsi),%xmm1
        movss (%rdi),%xmm2
        shufps $0,%xmm1,%xmm1
        shufps $0,%xmm2,%xmm2
        movaps %xmm1,nb211nf_krf(%rsp)
        movaps %xmm2,nb211nf_crf(%rsp)

        ## create constant floating-point factors on stack
        movl $0x3f000000,%eax   ## half in IEEE (hex)
        movl %eax,nb211nf_half(%rsp)
        movss nb211nf_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,nb211nf_half(%rsp)
        movaps %xmm3,nb211nf_three(%rsp)

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

        movq  nb211nf_charge(%rbp),%rdx
        movss (%rdx,%rbx,4),%xmm3
        movss 4(%rdx,%rbx,4),%xmm4
        movq nb211nf_p_facel(%rbp),%rsi
        movss (%rsi),%xmm0
        movss nb211nf_facel(%rsp),%xmm5
        mulss  %xmm5,%xmm3
        mulss  %xmm5,%xmm4

        shufps $0,%xmm3,%xmm3
        shufps $0,%xmm4,%xmm4
        movaps %xmm3,nb211nf_iqO(%rsp)
        movaps %xmm4,nb211nf_iqH(%rsp)

        movq  nb211nf_type(%rbp),%rdx
        movl  (%rdx,%rbx,4),%ecx
        shll  %ecx
        movq nb211nf_p_ntype(%rbp),%rdi
        imull (%rdi),%ecx     ## rcx = ntia = 2*ntype*type[ii0] 
        movl  %ecx,nb211nf_ntia(%rsp)

_nb_kernel211nf_x86_64_sse.nb211nf_threadloop: 
        movq  nb211nf_count(%rbp),%rsi            ## pointer to sync counter
        movl  (%rsi),%eax
_nb_kernel211nf_x86_64_sse.nb211nf_spinlock: 
        movl  %eax,%ebx                         ## ebx=*count=nn0
        addq  $1,%rbx                          ## rbx=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_kernel211nf_x86_64_sse.nb211nf_spinlock

        ## if(nn1>nri) nn1=nri
        movl nb211nf_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,nb211nf_n(%rsp)
        movl %ebx,nb211nf_nn1(%rsp)
        subl %eax,%ebx                          ## calc number of outer lists
        movl %eax,%esi                          ## copy n to esi
        jg  _nb_kernel211nf_x86_64_sse.nb211nf_outerstart
        jmp _nb_kernel211nf_x86_64_sse.nb211nf_end

_nb_kernel211nf_x86_64_sse.nb211nf_outerstart: 
        ## ebx contains number of outer iterations
        addl nb211nf_nouter(%rsp),%ebx
        movl %ebx,nb211nf_nouter(%rsp)

_nb_kernel211nf_x86_64_sse.nb211nf_outer: 
        movq  nb211nf_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,nb211nf_is3(%rsp)            ## store is3 

        movq  nb211nf_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  nb211nf_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  nb211nf_pos(%rbp),%rax      ## rax = base of pos[]  
        movl  %ebx,nb211nf_ii3(%rsp)

        addss (%rax,%rbx,4),%xmm3
        addss 4(%rax,%rbx,4),%xmm4
        addss 8(%rax,%rbx,4),%xmm5
        shufps $0,%xmm3,%xmm3
        shufps $0,%xmm4,%xmm4
        shufps $0,%xmm5,%xmm5
        movaps %xmm3,nb211nf_ixO(%rsp)
        movaps %xmm4,nb211nf_iyO(%rsp)
        movaps %xmm5,nb211nf_izO(%rsp)

        movss %xmm0,%xmm3
        movss %xmm1,%xmm4
        movss %xmm2,%xmm5
        addss 12(%rax,%rbx,4),%xmm0
        addss 16(%rax,%rbx,4),%xmm1
        addss 20(%rax,%rbx,4),%xmm2
        addss 24(%rax,%rbx,4),%xmm3
        addss 28(%rax,%rbx,4),%xmm4
        addss 32(%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,nb211nf_ixH1(%rsp)
        movaps %xmm1,nb211nf_iyH1(%rsp)
        movaps %xmm2,nb211nf_izH1(%rsp)
        movaps %xmm3,nb211nf_ixH2(%rsp)
        movaps %xmm4,nb211nf_iyH2(%rsp)
        movaps %xmm5,nb211nf_izH2(%rsp)

        ## clear vctot and i forces 
        xorps %xmm4,%xmm4
        movaps %xmm4,nb211nf_vctot(%rsp)
        movaps %xmm4,nb211nf_Vvdwtot(%rsp)

        movq  nb211nf_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  nb211nf_pos(%rbp),%rsi
        movq  nb211nf_jjnr(%rsp),%rax
        shll  $2,%ecx
        addq  %rcx,%rax
        movq  %rax,nb211nf_innerjjnr(%rsp)       ## pointer to jjnr[nj0] 
        movl  %edx,%ecx
        subl  $4,%edx
        addl  nb211nf_ninner(%rsp),%ecx
        movl  %ecx,nb211nf_ninner(%rsp)
        addl  $0,%edx
        movl  %edx,nb211nf_innerk(%rsp)      ## number of innerloop atoms