nb_kernel131_x86_64_sse2.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,044 行 · 第 1/4 页
S
2,044 行
movsd %xmm2,%xmm7 unpcklpd %xmm1,%xmm6 ## accumulate H1i forces in xmm0, xmm1, xmm2 movapd nb131_fixH1(%rsp),%xmm0 movapd nb131_fiyH1(%rsp),%xmm1 movapd nb131_fizH1(%rsp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addsd %xmm3,%xmm0 addsd %xmm4,%xmm1 addsd %xmm5,%xmm2 ## sum is in low xmm0-xmm2 ## increment i force movsd 24(%rdi,%rcx,8),%xmm3 movsd 32(%rdi,%rcx,8),%xmm4 movsd 40(%rdi,%rcx,8),%xmm5 subsd %xmm0,%xmm3 subsd %xmm1,%xmm4 subsd %xmm2,%xmm5 movsd %xmm3,24(%rdi,%rcx,8) movsd %xmm4,32(%rdi,%rcx,8) movsd %xmm5,40(%rdi,%rcx,8) ## accumulate force in xmm6/xmm7 for fshift addsd %xmm2,%xmm7 unpcklpd %xmm1,%xmm0 addpd %xmm0,%xmm6 ## accumulate H2i forces in xmm0, xmm1, xmm2 movapd nb131_fixH2(%rsp),%xmm0 movapd nb131_fiyH2(%rsp),%xmm1 movapd nb131_fizH2(%rsp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addsd %xmm3,%xmm0 addsd %xmm4,%xmm1 addsd %xmm5,%xmm2 ## sum is in low xmm0-xmm2 movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 ## increment i force movsd 48(%rdi,%rcx,8),%xmm3 movsd 56(%rdi,%rcx,8),%xmm4 movsd 64(%rdi,%rcx,8),%xmm5 subsd %xmm0,%xmm3 subsd %xmm1,%xmm4 subsd %xmm2,%xmm5 movsd %xmm3,48(%rdi,%rcx,8) movsd %xmm4,56(%rdi,%rcx,8) movsd %xmm5,64(%rdi,%rcx,8) ## accumulate force in xmm6/xmm7 for fshift addsd %xmm2,%xmm7 unpcklpd %xmm1,%xmm0 addpd %xmm0,%xmm6 ## increment fshift force movlpd (%rsi,%rdx,8),%xmm3 movhpd 8(%rsi,%rdx,8),%xmm3 movsd 16(%rsi,%rdx,8),%xmm4 subpd %xmm6,%xmm3 subsd %xmm7,%xmm4 movlpd %xmm3,(%rsi,%rdx,8) movhpd %xmm3,8(%rsi,%rdx,8) movsd %xmm4,16(%rsi,%rdx,8) ## get n from stack movl nb131_n(%rsp),%esi ## get group index for i particle movq nb131_gid(%rbp),%rdx ## base of gid[] movl (%rdx,%rsi,4),%edx ## ggid=gid[n] ## accumulate total potential energy and update it movapd nb131_vctot(%rsp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addsd %xmm6,%xmm7 ## low xmm7 has the sum now ## add earlier value from mem movq nb131_Vc(%rbp),%rax addsd (%rax,%rdx,8),%xmm7 ## move back to mem movsd %xmm7,(%rax,%rdx,8) ## accumulate total lj energy and update it movapd nb131_Vvdwtot(%rsp),%xmm7 ## accumulate movhlps %xmm7,%xmm6 addsd %xmm6,%xmm7 ## low xmm7 has the sum now ## add earlier value from mem movq nb131_Vvdw(%rbp),%rax addsd (%rax,%rdx,8),%xmm7 ## move back to mem movsd %xmm7,(%rax,%rdx,8) ## finish if last movl nb131_nn1(%rsp),%ecx ## esi already loaded with n incl %esi subl %esi,%ecx jz _nb_kernel131_x86_64_sse2.nb131_outerend ## not last, iterate outer loop once more! movl %esi,nb131_n(%rsp) jmp _nb_kernel131_x86_64_sse2.nb131_outer_nb_kernel131_x86_64_sse2.nb131_outerend: ## check if more outer neighborlists remain movl nb131_nri(%rsp),%ecx ## esi already loaded with n above subl %esi,%ecx jz _nb_kernel131_x86_64_sse2.nb131_end ## non-zero, do one more workunit jmp _nb_kernel131_x86_64_sse2.nb131_threadloop_nb_kernel131_x86_64_sse2.nb131_end: movl nb131_nouter(%rsp),%eax movl nb131_ninner(%rsp),%ebx movq nb131_outeriter(%rbp),%rcx movq nb131_inneriter(%rbp),%rdx movl %eax,(%rcx) movl %ebx,(%rdx) addq $920,%rsp emms pop %r15 pop %r14 pop %r13 pop %r12 pop %rbx pop %rbp ret.globl nb_kernel131nf_x86_64_sse2.globl _nb_kernel131nf_x86_64_sse2nb_kernel131nf_x86_64_sse2: _nb_kernel131nf_x86_64_sse2: ## Room for return address and rbp (16 bytes).set nb131nf_fshift, 16.set nb131nf_gid, 24.set nb131nf_pos, 32.set nb131nf_faction, 40.set nb131nf_charge, 48.set nb131nf_p_facel, 56.set nb131nf_argkrf, 64.set nb131nf_argcrf, 72.set nb131nf_Vc, 80.set nb131nf_type, 88.set nb131nf_p_ntype, 96.set nb131nf_vdwparam, 104.set nb131nf_Vvdw, 112.set nb131nf_p_tabscale, 120.set nb131nf_VFtab, 128.set nb131nf_invsqrta, 136.set nb131nf_dvda, 144.set nb131nf_p_gbtabscale, 152.set nb131nf_GBtab, 160.set nb131nf_p_nthreads, 168.set nb131nf_count, 176.set nb131nf_mtx, 184.set nb131nf_outeriter, 192.set nb131nf_inneriter, 200.set nb131nf_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb131nf_ixO, 0.set nb131nf_iyO, 16.set nb131nf_izO, 32.set nb131nf_ixH1, 48.set nb131nf_iyH1, 64.set nb131nf_izH1, 80.set nb131nf_ixH2, 96.set nb131nf_iyH2, 112.set nb131nf_izH2, 128.set nb131nf_iqO, 144.set nb131nf_iqH, 160.set nb131nf_qqO, 176.set nb131nf_qqH, 192.set nb131nf_c6, 208.set nb131nf_c12, 224.set nb131nf_vctot, 240.set nb131nf_Vvdwtot, 256.set nb131nf_half, 272.set nb131nf_three, 288.set nb131nf_rsqO, 304.set nb131nf_rinvH1, 320.set nb131nf_rinvH2, 336.set nb131nf_krsqO, 352.set nb131nf_krsqH1, 368.set nb131nf_krsqH2, 384.set nb131nf_tsc, 400.set nb131nf_facel, 416.set nb131nf_iinr, 424.set nb131nf_jindex, 432.set nb131nf_jjnr, 440.set nb131nf_shift, 448.set nb131nf_shiftvec, 456.set nb131nf_innerjjnr, 464.set nb131nf_nri, 472.set nb131nf_ntia, 476.set nb131nf_is3, 480.set nb131nf_ii3, 484.set nb131nf_innerk, 488.set nb131nf_n, 492.set nb131nf_nn1, 496.set nb131nf_nouter, 500.set nb131nf_ninner, 504 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $520,%rsp # # local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb131nf_nouter(%rsp) movl %eax,nb131nf_ninner(%rsp) movl (%rdi),%edi movl %edi,nb131nf_nri(%rsp) movq %rsi,nb131nf_iinr(%rsp) movq %rdx,nb131nf_jindex(%rsp) movq %rcx,nb131nf_jjnr(%rsp) movq %r8,nb131nf_shift(%rsp) movq %r9,nb131nf_shiftvec(%rsp) movq nb131nf_p_facel(%rbp),%rsi movsd (%rsi),%xmm0 movsd %xmm0,nb131nf_facel(%rsp) movq nb131nf_p_tabscale(%rbp),%rax movsd (%rax),%xmm3 shufpd $0,%xmm3,%xmm3 movapd %xmm3,nb131nf_tsc(%rsp) ## create constant floating-point factors on stack movl $0x00000000,%eax ## lower half of double half IEEE (hex) movl $0x3fe00000,%ebx movl %eax,nb131nf_half(%rsp) movl %ebx,nb131nf_half+4(%rsp) movsd nb131nf_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,nb131nf_half(%rsp) movapd %xmm3,nb131nf_three(%rsp) ## assume we have at least one i particle - start directly movq nb131nf_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx),%ebx ## ebx =ii movq nb131nf_charge(%rbp),%rdx movsd (%rdx,%rbx,8),%xmm3 movsd 8(%rdx,%rbx,8),%xmm4 movsd nb131nf_facel(%rsp),%xmm5 mulsd %xmm5,%xmm3 mulsd %xmm5,%xmm4 shufpd $0,%xmm3,%xmm3 shufpd $0,%xmm4,%xmm4 movapd %xmm3,nb131nf_iqO(%rsp) movapd %xmm4,nb131nf_iqH(%rsp) movq nb131nf_type(%rbp),%rdx movl (%rdx,%rbx,4),%ecx shll %ecx movq nb131nf_p_ntype(%rbp),%rdi imull (%rdi),%ecx ## rcx = ntia = 2*ntype*type[ii0] movl %ecx,nb131nf_ntia(%rsp)_nb_kernel131nf_x86_64_sse2.nb131nf_threadloop: movq nb131nf_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel131nf_x86_64_sse2.nb131nf_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_kernel131nf_x86_64_sse2.nb131nf_spinlock ## if(nn1>nri) nn1=nri movl nb131nf_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,nb131nf_n(%rsp) movl %ebx,nb131nf_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel131nf_x86_64_sse2.nb131nf_outerstart jmp _nb_kernel131nf_x86_64_sse2.nb131nf_end_nb_kernel131nf_x86_64_sse2.nb131nf_outerstart: ## ebx contains number of outer iterations addl nb131nf_nouter(%rsp),%ebx movl %ebx,nb131nf_nouter(%rsp)_nb_kernel131nf_x86_64_sse2.nb131nf_outer: movq nb131nf_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,nb131nf_is3(%rsp) ## store is3 movq nb131nf_shiftvec(%rsp),%rax ## eax = base of shiftvec[] movsd (%rax,%rbx,8),%xmm0 movsd 8(%rax,%rbx,8),%xmm1 movsd 16(%rax,%rbx,8),%xmm2 movq nb131nf_iinr(%rsp),%rcx ## ecx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movapd %xmm0,%xmm3 movapd %xmm1,%xmm4 movapd %xmm2,%xmm5 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb131nf_pos(%rbp),%rax ## eax = base of pos[] movl %ebx,nb131nf_ii3(%rsp) 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,nb131nf_ixO(%rsp) movapd %xmm4,nb131nf_iyO(%rsp) movapd %xmm5,nb131nf_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,nb131nf_ixH1(%rsp) movapd %xmm1,nb131nf_iyH1(%rsp) movapd %xmm2,nb131nf_izH1(%rsp) movapd %xmm3,nb131nf_ixH2(%rsp) movapd %xmm4,nb131nf_iyH2(%rsp) movapd %xmm5,nb131nf_izH2(%rsp) ## clear vctot xorpd %xmm4,%xmm4 movapd %xmm4,nb131nf_vctot(%rsp) movapd %xmm4,nb131nf_Vvdwtot(%rsp) movq nb131nf_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 nb131nf_pos(%rbp),%rsi movq nb131nf_faction(%rbp),%rdi movq nb131nf_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb131nf_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $2,%edx addl nb131nf_ninner(%rsp),%ecx movl %ecx,nb131nf_ninner(%rsp) addl $0,%edx movl %edx,nb131nf_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel131nf_x86_64_sse2.nb131nf_unroll_loop jmp _nb_kernel131nf_x86_64_sse2.nb131nf_checksingle_nb_kernel131nf_x86_64_sse2.nb131nf_unroll_loop: ## twice unrolled innerloop here movq nb131nf_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb131nf_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb131nf_charge(%rbp),%rsi ## base of charge[] movlpd (%rsi,%rax,8),%xmm3 movhpd (%rsi,%rbx,8),%xmm3 movapd %xmm3,%xmm4 mulpd nb131nf_iqO(%rsp),%xmm3 mulpd nb131nf_iqH(%rsp),%xmm4 movd %eax,%mm0 ## use mmx registers as temp storage movd %ebx,%mm1 movapd %xmm3,nb131nf_qqO(%rsp) movapd %xmm4,nb131nf_qqH(%rsp) movq nb131nf_type(%rbp),%rsi movl (%rsi,%rax,4),%eax movl (%rsi,%rbx,4),%ebx movq nb131nf_vdwparam(%rbp),%rsi shll %eax shll %ebx movl nb131nf_ntia(%rsp),%edi addl %edi,%eax addl %edi,%ebx movlpd (%rsi,%rax,8),%xmm6 ## c6a movlpd (%rsi,%rbx,8),%xmm7 ## c6b movhpd 8(%rsi,%rax,8),%xmm6 ## c6a c12a movhpd 8(%rsi,%rbx,8),%xmm7 ## c6b c12b movapd %xmm6,%xmm4 unpcklpd %xmm7,%xmm4 unpckhpd %xmm7,%xmm6 movd %mm0,%eax movd %mm1,%ebx movapd %xmm4,nb131nf_c6(%rsp) movapd %xmm6,nb131nf_c12(%rsp) movq nb131nf_pos(%rbp),%rsi ## base of pos[] lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## move two coordinates to xmm0-xmm2 movlpd (%rsi,%rax,8),%xmm0 movlpd 8(%rsi,%rax,8),%xmm1 movlpd 16(%rsi,%rax,8),%xmm2 movhpd (%rsi,%rbx,8),%xmm0 movhpd 8(%rsi,%rbx,8),%xmm1 movhpd 16(%rsi,%rbx,8),%xmm2 ## move ixO-izO to xmm4-xmm6 movapd nb131nf_ixO(%rsp),%xmm4 movapd nb131nf_iyO(%rsp),%xmm5 movapd nb131nf_izO(%rsp),%xmm6 ## calc dr subpd %xmm0,%xmm4 subpd %xmm1,%xmm5 subpd %xmm2,%xmm6 ## square it mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 mulpd %xmm6,%xmm6 addpd %xmm5,%xmm4 addpd %xmm6,%xmm4 movapd %xmm4,%xmm7 ## rsqO in xmm7 movapd %xmm7,nb131nf_rsqO(%rsp) ## move ixH1-izH1 to xmm4-xmm6 movapd nb131nf_ixH1(%rsp),%xmm4 movapd nb131nf_iyH1(%rsp),%xmm5 movapd nb131nf_izH1(%rsp),%xmm6 ## calc dr subpd %xmm0,%xmm4 subpd %xmm1,%xmm5 subpd %xmm2,%xmm6 ## square it mulpd %xmm4,%xmm4 mulpd %xmm5,%xmm5 mulpd %xmm6,%xmm6 addpd %xmm5,%xmm6 addpd %xmm4,%xmm6 ## rsqH1 in xmm6 ## move ixH2-izH2 to xmm3-xmm5 movapd nb131nf_ixH2(%rsp),%xmm3 movapd nb131nf_iyH2(%rsp),%xmm4 movapd nb131nf_izH2(%rsp),%xmm5 ## calc dr subpd %xmm0,%xmm3 subpd %xmm1,%xmm4 subpd %xmm2,%xmm5 ## square it ⌨️ 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?