nb_kernel102_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,099 行 · 第 1/5 页
S
2,099 行
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 nb102_fixH1(%rsp),%xmm0 movaps nb102_fiyH1(%rsp),%xmm1 movaps nb102_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 nb102_fixH2(%rsp),%xmm0 movaps nb102_fiyH2(%rsp),%xmm1 movaps nb102_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 nb102_n(%rsp),%esi ## get group index for i particle movq nb102_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movaps nb102_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 nb102_Vc(%rbp),%rax addss (%rax,%rdx,4),%xmm7 ## move back to mem movss %xmm7,(%rax,%rdx,4) ## finish if last movl nb102_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel102_x86_64_sse.nb102_outerend ## not last, iterate outer loop once more! movl %esi,nb102_n(%rsp) jmp _nb_kernel102_x86_64_sse.nb102_outer_nb_kernel102_x86_64_sse.nb102_outerend: ## check if more outer neighborlists remain movl nb102_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel102_x86_64_sse.nb102_end ## non-zero, do one more workunit jmp _nb_kernel102_x86_64_sse.nb102_threadloop_nb_kernel102_x86_64_sse.nb102_end: movl nb102_nouter(%rsp),%eax movl nb102_ninner(%rsp),%ebx movq nb102_outeriter(%rbp),%rcx movq nb102_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $1512,%rsp emms pop %r15 pop %r14 pop %r13 pop %r12 pop %rbx pop %rbp ret.globl nb_kernel102nf_x86_64_sse.globl _nb_kernel102nf_x86_64_ssenb_kernel102nf_x86_64_sse: _nb_kernel102nf_x86_64_sse: .set nb102nf_fshift, 16.set nb102nf_gid, 24.set nb102nf_pos, 32.set nb102nf_faction, 40.set nb102nf_charge, 48.set nb102nf_p_facel, 56.set nb102nf_argkrf, 64.set nb102nf_argcrf, 72.set nb102nf_Vc, 80.set nb102nf_type, 88.set nb102nf_p_ntype, 96.set nb102nf_vdwparam, 104.set nb102nf_Vvdw, 112.set nb102nf_p_tabscale, 120.set nb102nf_VFtab, 128.set nb102nf_invsqrta, 136.set nb102nf_dvda, 144.set nb102nf_p_gbtabscale, 152.set nb102nf_GBtab, 160.set nb102nf_p_nthreads, 168.set nb102nf_count, 176.set nb102nf_mtx, 184.set nb102nf_outeriter, 192.set nb102nf_inneriter, 200.set nb102nf_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb102nf_ixO, 0.set nb102nf_iyO, 16.set nb102nf_izO, 32.set nb102nf_ixH1, 48.set nb102nf_iyH1, 64.set nb102nf_izH1, 80.set nb102nf_ixH2, 96.set nb102nf_iyH2, 112.set nb102nf_izH2, 128.set nb102nf_jxO, 144.set nb102nf_jyO, 160.set nb102nf_jzO, 176.set nb102nf_jxH1, 192.set nb102nf_jyH1, 208.set nb102nf_jzH1, 224.set nb102nf_jxH2, 240.set nb102nf_jyH2, 256.set nb102nf_jzH2, 272.set nb102nf_qqOO, 288.set nb102nf_qqOH, 304.set nb102nf_qqHH, 320.set nb102nf_vctot, 336.set nb102nf_half, 352.set nb102nf_three, 368.set nb102nf_rsqOO, 384.set nb102nf_rsqOH1, 400.set nb102nf_rsqOH2, 416.set nb102nf_rsqH1O, 432.set nb102nf_rsqH1H1, 448.set nb102nf_rsqH1H2, 464.set nb102nf_rsqH2O, 480.set nb102nf_rsqH2H1, 496.set nb102nf_rsqH2H2, 512.set nb102nf_rinvOO, 528.set nb102nf_rinvOH1, 544.set nb102nf_rinvOH2, 560.set nb102nf_rinvH1O, 576.set nb102nf_rinvH1H1, 592.set nb102nf_rinvH1H2, 608.set nb102nf_rinvH2O, 624.set nb102nf_rinvH2H1, 640.set nb102nf_rinvH2H2, 656.set nb102nf_is3, 672.set nb102nf_ii3, 676.set nb102nf_nri, 680.set nb102nf_iinr, 688.set nb102nf_jindex, 696.set nb102nf_jjnr, 704.set nb102nf_shift, 712.set nb102nf_shiftvec, 720.set nb102nf_facel, 728.set nb102nf_innerjjnr, 736.set nb102nf_innerk, 744.set nb102nf_n, 748.set nb102nf_nn1, 752.set nb102nf_nouter, 756.set nb102nf_ninner, 760 push %rbp movq %rsp,%rbp push %rbx subq $776,%rsp emms ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb102nf_nouter(%rsp) movl %eax,nb102nf_ninner(%rsp) movl (%rdi),%edi movl %edi,nb102nf_nri(%rsp) movq %rsi,nb102nf_iinr(%rsp) movq %rdx,nb102nf_jindex(%rsp) movq %rcx,nb102nf_jjnr(%rsp) movq %r8,nb102nf_shift(%rsp) movq %r9,nb102nf_shiftvec(%rsp) movq nb102nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb102nf_facel(%rsp) ## assume we have at least one i particle - start directly movq nb102nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb102nf_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 movss %xmm3,%xmm4 movss 4(%rdx,%rbx,4),%xmm5 movq nb102nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb102nf_facel(%rsp),%xmm6 mulss %xmm3,%xmm3 mulss %xmm5,%xmm4 mulss %xmm5,%xmm5 mulss %xmm6,%xmm3 mulss %xmm6,%xmm4 mulss %xmm6,%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movaps %xmm3,nb102nf_qqOO(%rsp) movaps %xmm4,nb102nf_qqOH(%rsp) movaps %xmm5,nb102nf_qqHH(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb102nf_half(%rsp) movss nb102nf_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,nb102nf_half(%rsp) movaps %xmm3,nb102nf_three(%rsp)_nb_kernel102nf_x86_64_sse.nb102nf_threadloop: movq nb102nf_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel102nf_x86_64_sse.nb102nf_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_kernel102nf_x86_64_sse.nb102nf_spinlock ## if(nn1>nri) nn1=nri movl nb102nf_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,nb102nf_n(%rsp) movl %ebx,nb102nf_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel102nf_x86_64_sse.nb102nf_outerstart jmp _nb_kernel102nf_x86_64_sse.nb102nf_end_nb_kernel102nf_x86_64_sse.nb102nf_outerstart: ## ebx contains number of outer iterations addl nb102nf_nouter(%rsp),%ebx movl %ebx,nb102nf_nouter(%rsp)_nb_kernel102nf_x86_64_sse.nb102nf_outer: movq nb102nf_shift(%rsp),%rax ## rax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## ebx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb102nf_is3(%rsp) ## store is3 movq nb102nf_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 nb102nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb102nf_pos(%rbp),%rax ## rax = base of pos[] movl %ebx,nb102nf_ii3(%rsp) movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 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,nb102nf_ixO(%rsp) movaps %xmm4,nb102nf_iyO(%rsp) movaps %xmm5,nb102nf_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,nb102nf_ixH1(%rsp) movaps %xmm1,nb102nf_iyH1(%rsp) movaps %xmm2,nb102nf_izH1(%rsp) movaps %xmm3,nb102nf_ixH2(%rsp) movaps %xmm4,nb102nf_iyH2(%rsp) movaps %xmm5,nb102nf_izH2(%rsp) ## clear vctot xorps %xmm4,%xmm4 movaps %xmm4,nb102nf_vctot(%rsp) movq nb102nf_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 nb102nf_pos(%rbp),%rsi⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?