nb_kernel133_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,130 行 · 第 1/5 页
S
2,130 行
movl nb133_ninner(%rsp),%ebx movq nb133_outeriter(%rbp),%rcx movq nb133_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_kernel133nf_x86_64_sse.globl _nb_kernel133nf_x86_64_ssenb_kernel133nf_x86_64_sse: _nb_kernel133nf_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb133nf_fshift, 16.set nb133nf_gid, 24.set nb133nf_pos, 32.set nb133nf_faction, 40.set nb133nf_charge, 48.set nb133nf_p_facel, 56.set nb133nf_argkrf, 64.set nb133nf_argcrf, 72.set nb133nf_Vc, 80.set nb133nf_type, 88.set nb133nf_p_ntype, 96.set nb133nf_vdwparam, 104.set nb133nf_Vvdw, 112.set nb133nf_p_tabscale, 120.set nb133nf_VFtab, 128.set nb133nf_invsqrta, 136.set nb133nf_dvda, 144.set nb133nf_p_gbtabscale, 152.set nb133nf_GBtab, 160.set nb133nf_p_nthreads, 168.set nb133nf_count, 176.set nb133nf_mtx, 184.set nb133nf_outeriter, 192.set nb133nf_inneriter, 200.set nb133nf_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb133nf_ixO, 0.set nb133nf_iyO, 16.set nb133nf_izO, 32.set nb133nf_ixH1, 48.set nb133nf_iyH1, 64.set nb133nf_izH1, 80.set nb133nf_ixH2, 96.set nb133nf_iyH2, 112.set nb133nf_izH2, 128.set nb133nf_ixM, 144.set nb133nf_iyM, 160.set nb133nf_izM, 176.set nb133nf_iqM, 192.set nb133nf_iqH, 208.set nb133nf_qqM, 224.set nb133nf_qqH, 240.set nb133nf_rinvH1, 256.set nb133nf_rinvH2, 272.set nb133nf_rinvM, 288.set nb133nf_tsc, 304.set nb133nf_c6, 320.set nb133nf_c12, 336.set nb133nf_krf, 352.set nb133nf_crf, 368.set nb133nf_krsqH1, 384.set nb133nf_krsqH2, 400.set nb133nf_krsqM, 416.set nb133nf_vctot, 432.set nb133nf_Vvdwtot, 448.set nb133nf_half, 464.set nb133nf_three, 480.set nb133nf_nri, 496.set nb133nf_iinr, 504.set nb133nf_jindex, 512.set nb133nf_jjnr, 520.set nb133nf_shift, 528.set nb133nf_shiftvec, 536.set nb133nf_facel, 544.set nb133nf_innerjjnr, 552.set nb133nf_is3, 560.set nb133nf_ii3, 564.set nb133nf_ntia, 568.set nb133nf_innerk, 572.set nb133nf_n, 576.set nb133nf_nn1, 580.set nb133nf_nouter, 584.set nb133nf_ninner, 588 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $600,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb133nf_nouter(%rsp) movl %eax,nb133nf_ninner(%rsp) movl (%rdi),%edi movl %edi,nb133nf_nri(%rsp) movq %rsi,nb133nf_iinr(%rsp) movq %rdx,nb133nf_jindex(%rsp) movq %rcx,nb133nf_jjnr(%rsp) movq %r8,nb133nf_shift(%rsp) movq %r9,nb133nf_shiftvec(%rsp) movq nb133nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb133nf_facel(%rsp) movq nb133nf_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb133nf_tsc(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb133nf_half(%rsp) movss nb133nf_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,nb133nf_half(%rsp) movaps %xmm3,nb133nf_three(%rsp) ## assume we have at least one i particle - start directly movq nb133nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb133nf_charge(%rbp),%rdx movss 4(%rdx,%rbx,4),%xmm4 movss 12(%rdx,%rbx,4),%xmm3 movq nb133nf_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss nb133nf_facel(%rsp),%xmm5 mulss %xmm5,%xmm3 mulss %xmm5,%xmm4 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 movaps %xmm3,nb133nf_iqM(%rsp) movaps %xmm4,nb133nf_iqH(%rsp) movq nb133nf_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb133nf_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb133nf_ntia(%rsp)_nb_kernel133nf_x86_64_sse.nb133nf_threadloop: movq nb133nf_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel133nf_x86_64_sse.nb133nf_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_kernel133nf_x86_64_sse.nb133nf_spinlock ## if(nn1>nri) nn1=nri movl nb133nf_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,nb133nf_n(%rsp) movl %ebx,nb133nf_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel133nf_x86_64_sse.nb133nf_outerstart jmp _nb_kernel133nf_x86_64_sse.nb133nf_end_nb_kernel133nf_x86_64_sse.nb133nf_outerstart: ## ebx contains number of outer iterations addl nb133nf_nouter(%rsp),%ebx movl %ebx,nb133nf_nouter(%rsp)_nb_kernel133nf_x86_64_sse.nb133nf_outer: movq nb133nf_shift(%rsp),%rax ## eax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## ebx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb133nf_is3(%rsp) ## store is3 movq nb133nf_shiftvec(%rsp),%rax ## eax = base of shiftvec[] movss (%rax,%rbx,4),%xmm0 movss 4(%rax,%rbx,4),%xmm1 movss 8(%rax,%rbx,4),%xmm2 movq nb133nf_iinr(%rsp),%rcx ## ecx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb133nf_pos(%rbp),%rax ## eax = base of pos[] movl %ebx,nb133nf_ii3(%rsp) addss (%rax,%rbx,4),%xmm3 ## ox addss 4(%rax,%rbx,4),%xmm4 ## oy addss 8(%rax,%rbx,4),%xmm5 ## oz addss 12(%rax,%rbx,4),%xmm6 ## h1x addss 16(%rax,%rbx,4),%xmm7 ## h1y shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 shufps $0,%xmm6,%xmm6 shufps $0,%xmm7,%xmm7 movaps %xmm3,nb133nf_ixO(%rsp) movaps %xmm4,nb133nf_iyO(%rsp) movaps %xmm5,nb133nf_izO(%rsp) movaps %xmm6,nb133nf_ixH1(%rsp) movaps %xmm7,nb133nf_iyH1(%rsp) movss %xmm2,%xmm6 movss %xmm0,%xmm3 movss %xmm1,%xmm4 movss %xmm2,%xmm5 addss 20(%rax,%rbx,4),%xmm6 ## h1z addss 24(%rax,%rbx,4),%xmm0 ## h2x addss 28(%rax,%rbx,4),%xmm1 ## h2y addss 32(%rax,%rbx,4),%xmm2 ## h2z addss 36(%rax,%rbx,4),%xmm3 ## mx addss 40(%rax,%rbx,4),%xmm4 ## my addss 44(%rax,%rbx,4),%xmm5 ## mz shufps $0,%xmm6,%xmm6 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 %xmm6,nb133nf_izH1(%rsp) movaps %xmm0,nb133nf_ixH2(%rsp) movaps %xmm1,nb133nf_iyH2(%rsp) movaps %xmm2,nb133nf_izH2(%rsp) movaps %xmm3,nb133nf_ixM(%rsp) movaps %xmm4,nb133nf_iyM(%rsp) movaps %xmm5,nb133nf_izM(%rsp) ## clear vctot xorps %xmm4,%xmm4 movaps %xmm4,nb133nf_vctot(%rsp) movaps %xmm4,nb133nf_Vvdwtot(%rsp) movq nb133nf_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 nb133nf_pos(%rbp),%rsi movq nb133nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb133nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb133nf_ninner(%rsp),%ecx movl %ecx,nb133nf_ninner(%rsp) addl $0,%edx movl %edx,nb133nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel133nf_x86_64_sse.nb133nf_unroll_loop jmp _nb_kernel133nf_x86_64_sse.nb133nf_odd_inner_nb_kernel133nf_x86_64_sse.nb133nf_unroll_loop: ## quad-unroll innerloop here movq nb133nf_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,nb133nf_innerjjnr(%rsp) ## advance pointer (unrolled 4) movq nb133nf_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 ## constant 10001000 ;# all charges in xmm3 movaps %xmm3,%xmm4 ## and in xmm4 mulps nb133nf_iqM(%rsp),%xmm3 mulps nb133nf_iqH(%rsp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movd %ecx,%mm2 movd %edx,%mm3 movaps %xmm3,nb133nf_qqM(%rsp) movaps %xmm4,nb133nf_qqH(%rsp) movq nb133nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movl (%rsi,%rbx,4),%ebx movl (%rsi,%rcx,4),%ecx movl (%rsi,%rdx,4),%edx movq nb133nf_vdwparam(%rbp),%rsi shll %eax shll %ebx shll %ecx shll %edx movl nb133nf_ntia(%rsp),%edi addl %edi,%eax addl %edi,%ebx addl %edi,%ecx addl %edi,%edx movlps (%rsi,%rax,4),%xmm6 movlps (%rsi,%rcx,4),%xmm7 movhps (%rsi,%rbx,4),%xmm6 movhps (%rsi,%rdx,4),%xmm7 movaps %xmm6,%xmm4 shufps $136,%xmm7,%xmm4 ## constant 10001000 shufps $221,%xmm7,%xmm6 ## constant 11011101 movd %mm0,%eax movd %mm1,%ebx movd %mm2,%ecx movd %mm3,%edx movaps %xmm4,nb133nf_c6(%rsp) movaps %xmm6,nb133nf_c12(%rsp) movq nb133nf_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 shufps $0,%xmm1,%xmm6 movaps %xmm4,%xmm0 movaps %xmm4,%xmm1 shufps $136,%xmm6,%xmm2 ## constant 10001000 shufps $136,%xmm5,%xmm0 ## constant 10001000 shufps $221,%xmm5,%xmm1 ## constant 11011101 ## move ixO-izO to xmm4-xmm6 movaps nb133nf_ixO(%rsp),%xmm4 movaps nb133nf_iyO(%rsp),%xmm5 movaps nb133nf_izO(%rsp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5 subps %xmm2,%xmm6 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 movaps %xmm4,%xmm7 ## rsqO in xmm7 ## move ixH1-izH1 to xmm4-xmm6 movaps nb133nf_ixH1(%rsp),%xmm4 movaps nb133nf_iyH1(%rsp),%xmm5 movaps nb133nf_izH1(%rsp),%xmm6 ## calc dr subps %xmm0,%xmm4 subps %xmm1,%xmm5⌨️ 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?