nb_kernel102_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,225 行 · 第 1/5 页
S
2,225 行
.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 sse2 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 emms push %r12 push %r13 push %r14 push %r15 subq $792,%rsp ## local variable stack space (n*16+8) ## 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 movsd (%rsi),%xmm0 movsd %xmm0,nb102nf_facel(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb102nf_half(%rsp) movl %ebx,nb102nf_half+4(%rsp) movsd nb102nf_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,nb102nf_half(%rsp) movapd %xmm3,nb102nf_three(%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 movsd (%rdx,%rbx,8),%xmm3 ## qO movsd %xmm3,%xmm4 ## qO movsd 8(%rdx,%rbx,8),%xmm5 ## qH movq nb102nf_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd nb102nf_facel(%rsp),%xmm6 ## facel mulsd %xmm3,%xmm3 ## qO*qO mulsd %xmm5,%xmm4 ## qO*qH mulsd %xmm5,%xmm5 ## qH*qH mulsd %xmm6,%xmm3 mulsd %xmm6,%xmm4 mulsd %xmm6,%xmm5 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 movapd %xmm3,nb102nf_qqOO(%rsp) movapd %xmm4,nb102nf_qqOH(%rsp) movapd %xmm5,nb102nf_qqHH(%rsp)_nb_kernel102nf_x86_64_sse2.nb102nf_threadloop: movq nb102nf_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel102nf_x86_64_sse2.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_sse2.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_sse2.nb102nf_outerstart jmp _nb_kernel102nf_x86_64_sse2.nb102nf_end_nb_kernel102nf_x86_64_sse2.nb102nf_outerstart: ## ebx contains number of outer iterations addl nb102nf_nouter(%rsp),%ebx movl %ebx,nb102nf_nouter(%rsp)_nb_kernel102nf_x86_64_sse2.nb102nf_outer: movq nb102nf_shift(%rsp),%rax ## rax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## rbx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movq nb102nf_shiftvec(%rsp),%rax ## rax = base of shiftvec[] movsd (%rax,%rbx,8),%xmm0 movsd 8(%rax,%rbx,8),%xmm1 movsd 16(%rax,%rbx,8),%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) movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 addsd (%rax,%rbx,8),%xmm3 addsd 8(%rax,%rbx,8),%xmm4 addsd 16(%rax,%rbx,8),%xmm5 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 movapd %xmm3,nb102nf_ixO(%rsp) movapd %xmm4,nb102nf_iyO(%rsp) movapd %xmm5,nb102nf_izO(%rsp) movsd %xmm0,%xmm3 movsd %xmm1,%xmm4 movsd %xmm2,%xmm5 addsd 24(%rax,%rbx,8),%xmm0 addsd 32(%rax,%rbx,8),%xmm1 addsd 40(%rax,%rbx,8),%xmm2 addsd 48(%rax,%rbx,8),%xmm3 addsd 56(%rax,%rbx,8),%xmm4 addsd 64(%rax,%rbx,8),%xmm5 shufpd $0,%xmm0,%xmm0 shufpd $0,%xmm1,%xmm1 shufpd $0,%xmm2,%xmm2 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 shufpd $0,%xmm5,%xmm5 movapd %xmm0,nb102nf_ixH1(%rsp) movapd %xmm1,nb102nf_iyH1(%rsp) movapd %xmm2,nb102nf_izH1(%rsp) movapd %xmm3,nb102nf_ixH2(%rsp) movapd %xmm4,nb102nf_iyH2(%rsp) movapd %xmm5,nb102nf_izH2(%rsp) ## clear vctot xorpd %xmm4,%xmm4 movapd %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 movq nb102nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb102nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb102nf_ninner(%rsp),%ecx movl %ecx,nb102nf_ninner(%rsp) addl $0,%edx movl %edx,nb102nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel102nf_x86_64_sse2.nb102nf_unroll_loop jmp _nb_kernel102nf_x86_64_sse2.nb102nf_checksingle_nb_kernel102nf_x86_64_sse2.nb102nf_unroll_loop: ## twice unrolled innerloop here movq nb102nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb102nf_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb102nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## move j coordinates to local temp variables movlpd (%rsi,%rax,8),%xmm2 movlpd 8(%rsi,%rax,8),%xmm3 movlpd 16(%rsi,%rax,8),%xmm4 movlpd 24(%rsi,%rax,8),%xmm5 movlpd 32(%rsi,%rax,8),%xmm6 movlpd 40(%rsi,%rax,8),%xmm7 movhpd (%rsi,%rbx,8),%xmm2 movhpd 8(%rsi,%rbx,8),%xmm3 movhpd 16(%rsi,%rbx,8),%xmm4 movhpd 24(%rsi,%rbx,8),%xmm5 movhpd 32(%rsi,%rbx,8),%xmm6 movhpd 40(%rsi,%rbx,8),%xmm7 movapd %xmm2,nb102nf_jxO(%rsp) movapd %xmm3,nb102nf_jyO(%rsp) movapd %xmm4,nb102nf_jzO(%rsp) movapd %xmm5,nb102nf_jxH1(%rsp) movapd %xmm6,nb102nf_jyH1(%rsp) movapd %xmm7,nb102nf_jzH1(%rsp) movlpd 48(%rsi,%rax,8),%xmm2 movlpd 56(%rsi,%rax,8),%xmm3 movlpd 64(%rsi,%rax,8),%xmm4 movhpd 48(%rsi,%rbx,8),%xmm2 movhpd 56(%rsi,%rbx,8),%xmm3 movhpd 64(%rsi,%rbx,8),%xmm4 movapd %xmm2,nb102nf_jxH2(%rsp) movapd %xmm3,nb102nf_jyH2(%rsp) movapd %xmm4,nb102nf_jzH2(%rsp) movapd nb102nf_ixO(%rsp),%xmm0 movapd nb102nf_iyO(%rsp),%xmm1 movapd nb102nf_izO(%rsp),%xmm2 movapd nb102nf_ixO(%rsp),%xmm3 movapd nb102nf_iyO(%rsp),%xmm4 movapd nb102nf_izO(%rsp),%xmm5 subpd nb102nf_jxO(%rsp),%xmm0 subpd nb102nf_jyO(%rsp),%xmm1 subpd nb102nf_jzO(%rsp),%xmm2 subpd nb102nf_jxH1(%rsp),%xmm3 subpd nb102nf_jyH1(%rsp),%xmm4 subpd nb102nf_jzH1(%rsp),%xmm5 mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 addpd %xmm1,%xmm0 addpd %xmm2,%xmm0 addpd %xmm4,%xmm3 addpd %xmm5,%xmm3 movapd %xmm0,nb102nf_rsqOO(%rsp) movapd %xmm3,nb102nf_rsqOH1(%rsp) movapd nb102nf_ixO(%rsp),%xmm0 movapd nb102nf_iyO(%rsp),%xmm1 movapd nb102nf_izO(%rsp),%xmm2 movapd nb102nf_ixH1(%rsp),%xmm3 movapd nb102nf_iyH1(%rsp),%xmm4 movapd nb102nf_izH1(%rsp),%xmm5 subpd nb102nf_jxH2(%rsp),%xmm0 subpd nb102nf_jyH2(%rsp),%xmm1 subpd nb102nf_jzH2(%rsp),%xmm2 subpd nb102nf_jxO(%rsp),%xmm3 subpd nb102nf_jyO(%rsp),%xmm4 subpd nb102nf_jzO(%rsp),%xmm5 mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 addpd %xmm1,%xmm0 addpd %xmm2,%xmm0 addpd %xmm4,%xmm3 addpd %xmm5,%xmm3 movapd %xmm0,nb102nf_rsqOH2(%rsp) movapd %xmm3,nb102nf_rsqH1O(%rsp) movapd nb102nf_ixH1(%rsp),%xmm0 movapd nb102nf_iyH1(%rsp),%xmm1 movapd nb102nf_izH1(%rsp),%xmm2 movapd nb102nf_ixH1(%rsp),%xmm3 movapd nb102nf_iyH1(%rsp),%xmm4 movapd nb102nf_izH1(%rsp),%xmm5 subpd nb102nf_jxH1(%rsp),%xmm0 subpd nb102nf_jyH1(%rsp),%xmm1 subpd nb102nf_jzH1(%rsp),%xmm2 subpd nb102nf_jxH2(%rsp),%xmm3 subpd nb102nf_jyH2(%rsp),%xmm4 subpd nb102nf_jzH2(%rsp),%xmm5 mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 addpd %xmm1,%xmm0 addpd %xmm2,%xmm0 addpd %xmm4,%xmm3 addpd %xmm5,%xmm3 movapd %xmm0,nb102nf_rsqH1H1(%rsp) movapd %xmm3,nb102nf_rsqH1H2(%rsp) movapd nb102nf_ixH2(%rsp),%xmm0 movapd nb102nf_iyH2(%rsp),%xmm1 movapd nb102nf_izH2(%rsp),%xmm2 movapd nb102nf_ixH2(%rsp),%xmm3 movapd nb102nf_iyH2(%rsp),%xmm4 movapd nb102nf_izH2(%rsp),%xmm5 subpd nb102nf_jxO(%rsp),%xmm0 subpd nb102nf_jyO(%rsp),%xmm1 subpd nb102nf_jzO(%rsp),%xmm2 subpd nb102nf_jxH1(%rsp),%xmm3 subpd nb102nf_jyH1(%rsp),%xmm4 subpd nb102nf_jzH1(%rsp),%xmm5 mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 mulpd %xmm3,%xmm3 mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 addpd %xmm1,%xmm0 addpd %xmm2,%xmm0 addpd %xmm3,%xmm4 addpd %xmm5,%xmm4 movapd %xmm0,nb102nf_rsqH2O(%rsp) movapd %xmm4,nb102nf_rsqH2H1(%rsp) movapd nb102nf_ixH2(%rsp),%xmm0 movapd nb102nf_iyH2(%rsp),%xmm1 movapd nb102nf_izH2(%rsp),%xmm2 subpd nb102nf_jxH2(%rsp),%xmm0 subpd nb102nf_jyH2(%rsp),%xmm1 subpd nb102nf_jzH2(%rsp),%xmm2 mulpd %xmm0,%xmm0 mulpd %xmm1,%xmm1 mulpd %xmm2,%xmm2 addpd %xmm1,%xmm0 addpd %xmm2,%xmm0 movapd %xmm0,nb102nf_rsqH2H2(%rsp) ## start doing invsqrt use rsq values in xmm0 (h2h2) , xmm4 (h2h1) cvtpd2ps %xmm0,%xmm1 cvtpd2ps %xmm4,%xmm5 rsqrtps %xmm1,%xmm1 rsqrtps %xmm5,%xmm5 cvtps2pd %xmm1,%xmm1 cvtps2pd %xmm5,%xmm5 movapd %xmm1,%xmm2 ## copy of luA movapd %xmm5,%xmm6 ## copy of luB mulpd %xmm1,%xmm1 ## luA*luA mulpd %xmm5,%xmm5 ## luB*luB movapd nb102nf_three(%rsp),%xmm3 mulpd %xmm0,%xmm1 ## rsqA*luA*luA mulpd %xmm4,%xmm5 ## rsqB*luB*luB movapd %xmm3,%xmm7 subpd %xmm1,%xmm3 ## 3-rsqA*luA*luA subpd %xmm5,%xmm7 ## 3-rsqB*luB*luB mulpd %xmm2,%xmm3 ## luA*(3-rsqA*luA*luA) mulpd %xmm6,%xmm7 ## luB*(3-rsqB*luB*luB) mulpd nb102nf_half(%rsp),%xmm3 ## iter1 mulpd nb102nf_half(%rsp),%xmm7 ## iter1 movapd %xmm3,%xmm2 ## copy of luA movapd %xmm7,%xmm6 ## copy of luB mulpd %xmm3,%xmm3 ## luA*luA mulpd %xmm7,%xmm7 ## luB*luB movapd nb102nf_three(%rsp),%xmm1 mulpd %xmm0,%xmm3 ## rsqA*luA*luA mulpd %xmm4,%xmm7 ## rsqB*luB*luB movapd %xmm1,%xmm5 subpd %xmm3,%xmm1 ## 3-rsqA*luA*luA subpd %xmm7,%xmm5 ## 3-rsqB*luB*luB mulpd %xmm2,%xmm1 ## luA*(3-rsqA*luA*luA) mulpd %xmm6,%xmm5 ## luB*(3-rsqB*luB*luB) mulpd nb102nf_half(%rsp),%xmm1 ## rinv mulpd nb102nf_half(%rsp),%xmm5 ## rinv movapd %xmm1,nb102nf_rinvH2H2(%rsp) movapd %xmm5,nb102nf_rinvH2H1(%rsp) movapd nb102nf_rsqOO(%rsp),%xmm0 ⌨️ 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?