nb_kernel213_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,195 行 · 第 1/4 页
S
2,195 行
mulsd xmm5, [rsp + nb213_qqM] ;# vcoul mulsd xmm7, [rsp + nb213_qqM] mulsd xmm4, xmm7 ;# total fsH2 in xmm4 addsd xmm5, [rsp + nb213_vctot] movapd xmm0, [rsp + nb213_dxM] movapd xmm1, [rsp + nb213_dyM] movapd xmm2, [rsp + nb213_dzM] movsd [rsp + nb213_vctot], xmm5 mulsd xmm0, xmm4 mulsd xmm1, xmm4 mulsd xmm2, xmm4 ;# update M forces movapd xmm3, [rsp + nb213_fixM] movapd xmm4, [rsp + nb213_fiyM] movapd xmm7, [rsp + nb213_fizM] addsd xmm3, xmm0 addsd xmm4, xmm1 addsd xmm7, xmm2 movsd [rsp + nb213_fixM], xmm3 movsd [rsp + nb213_fiyM], xmm4 movsd [rsp + nb213_fizM], xmm7 mov rdi, [rbp + nb213_faction] ;# update j forces addsd xmm0, [rsp + nb213_fjx] addsd xmm1, [rsp + nb213_fjy] addsd xmm2, [rsp + nb213_fjz] movlpd xmm3, [rdi + rax*8] movlpd xmm4, [rdi + rax*8 + 8] movlpd xmm5, [rdi + rax*8 + 16] addsd xmm3, xmm0 addsd xmm4, xmm1 addsd xmm5, xmm2 movlpd [rdi + rax*8], xmm3 movlpd [rdi + rax*8 + 8], xmm4 movlpd [rdi + rax*8 + 16], xmm5.nb213_updateouterdata: mov ecx, [rsp + nb213_ii3] mov rdi, [rbp + nb213_faction] mov rsi, [rbp + nb213_fshift] mov edx, [rsp + nb213_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movapd xmm0, [rsp + nb213_fixO] movapd xmm1, [rsp + nb213_fiyO] movapd xmm2, [rsp + nb213_fizO] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addsd xmm0, xmm3 addsd xmm1, xmm4 addsd xmm2, xmm5 ;# sum is in low xmm0-xmm2 ;# increment i force movsd xmm3, [rdi + rcx*8] movsd xmm4, [rdi + rcx*8 + 8] movsd xmm5, [rdi + rcx*8 + 16] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 movsd [rdi + rcx*8], xmm3 movsd [rdi + rcx*8 + 8], xmm4 movsd [rdi + rcx*8 + 16], xmm5 ;# accumulate force in xmm6/xmm7 for fshift movapd xmm6, xmm0 movsd xmm7, xmm2 unpcklpd xmm6,xmm1 ;# accumulate H1i forces in xmm0, xmm1, xmm2 movapd xmm0, [rsp + nb213_fixH1] movapd xmm1, [rsp + nb213_fiyH1] movapd xmm2, [rsp + nb213_fizH1] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addsd xmm0, xmm3 addsd xmm1, xmm4 addsd xmm2, xmm5 ;# sum is in low xmm0-xmm2 ;# increment i force movsd xmm3, [rdi + rcx*8 + 24] movsd xmm4, [rdi + rcx*8 + 32] movsd xmm5, [rdi + rcx*8 + 40] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 movsd [rdi + rcx*8 + 24], xmm3 movsd [rdi + rcx*8 + 32], xmm4 movsd [rdi + rcx*8 + 40], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addsd xmm7, xmm2 unpcklpd xmm0, xmm1 addpd xmm6, xmm0 ;# accumulate H2i forces in xmm0, xmm1, xmm2 movapd xmm0, [rsp + nb213_fixH2] movapd xmm1, [rsp + nb213_fiyH2] movapd xmm2, [rsp + nb213_fizH2] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addsd xmm0, xmm3 addsd xmm1, xmm4 addsd xmm2, xmm5 ;# sum is in low xmm0-xmm2 ;# increment i force movsd xmm3, [rdi + rcx*8 + 48] movsd xmm4, [rdi + rcx*8 + 56] movsd xmm5, [rdi + rcx*8 + 64] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 movsd [rdi + rcx*8 + 48], xmm3 movsd [rdi + rcx*8 + 56], xmm4 movsd [rdi + rcx*8 + 64], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addsd xmm7, xmm2 unpcklpd xmm0, xmm1 addpd xmm6, xmm0 ;# accumulate Mi forces in xmm0, xmm1, xmm2 movapd xmm0, [rsp + nb213_fixM] movapd xmm1, [rsp + nb213_fiyM] movapd xmm2, [rsp + nb213_fizM] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addsd xmm0, xmm3 addsd xmm1, xmm4 addsd xmm2, xmm5 ;# sum is in low xmm0-xmm2 ;# increment i force movsd xmm3, [rdi + rcx*8 + 72] movsd xmm4, [rdi + rcx*8 + 80] movsd xmm5, [rdi + rcx*8 + 88] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 movsd [rdi + rcx*8 + 72], xmm3 movsd [rdi + rcx*8 + 80], xmm4 movsd [rdi + rcx*8 + 88], xmm5 ;# accumulate force in xmm6/xmm7 for fshift addsd xmm7, xmm2 unpcklpd xmm0, xmm1 addpd xmm6, xmm0 ;# increment fshift force movlpd xmm3, [rsi + rdx*8] movhpd xmm3, [rsi + rdx*8 + 8] movsd xmm4, [rsi + rdx*8 + 16] subpd xmm3, xmm6 subsd xmm4, xmm7 movlpd [rsi + rdx*8], xmm3 movhpd [rsi + rdx*8 + 8], xmm3 movsd [rsi + rdx*8 + 16], xmm4 ;# get n from stack mov esi, [rsp + nb213_n] ;# get group index for i particle mov rdx, [rbp + nb213_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movapd xmm7, [rsp + nb213_vctot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# low xmm7 has the sum now ;# add earlier value from mem mov rax, [rbp + nb213_Vc] addsd xmm7, [rax + rdx*8] ;# move back to mem movsd [rax + rdx*8], xmm7 ;# accumulate total lj energy and update it movapd xmm7, [rsp + nb213_Vvdwtot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# low xmm7 has the sum now ;# add earlier value from mem mov rax, [rbp + nb213_Vvdw] addsd xmm7, [rax + rdx*8] ;# move back to mem movsd [rax + rdx*8], xmm7 ;# finish if last mov ecx, [rsp + nb213_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb213_outerend ;# not last, iterate outer loop once more! mov [rsp + nb213_n], esi jmp .nb213_outer.nb213_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb213_nri] ;# esi already loaded with n above sub ecx, esi jz .nb213_end ;# non-zero, do one more workunit jmp .nb213_threadloop.nb213_end: mov eax, [rsp + nb213_nouter] mov ebx, [rsp + nb213_ninner] mov rcx, [rbp + nb213_outeriter] mov rdx, [rbp + nb213_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 1064 emms pop r15 pop r14 pop r13 pop r12 pop rbx pop rbp ret.globl nb_kernel213nf_x86_64_sse2.globl _nb_kernel213nf_x86_64_sse2nb_kernel213nf_x86_64_sse2: _nb_kernel213nf_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb213nf_fshift, 16.equiv nb213nf_gid, 24.equiv nb213nf_pos, 32.equiv nb213nf_faction, 40.equiv nb213nf_charge, 48.equiv nb213nf_p_facel, 56.equiv nb213nf_argkrf, 64.equiv nb213nf_argcrf, 72.equiv nb213nf_Vc, 80.equiv nb213nf_type, 88.equiv nb213nf_p_ntype, 96.equiv nb213nf_vdwparam, 104.equiv nb213nf_Vvdw, 112.equiv nb213nf_p_tabscale, 120.equiv nb213nf_VFtab, 128.equiv nb213nf_invsqrta, 136.equiv nb213nf_dvda, 144.equiv nb213nf_p_gbtabscale, 152.equiv nb213nf_GBtab, 160.equiv nb213nf_p_nthreads, 168.equiv nb213nf_count, 176.equiv nb213nf_mtx, 184.equiv nb213nf_outeriter, 192.equiv nb213nf_inneriter, 200.equiv nb213nf_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse2 use .equiv nb213nf_ixO, 0.equiv nb213nf_iyO, 16.equiv nb213nf_izO, 32.equiv nb213nf_ixH1, 48.equiv nb213nf_iyH1, 64.equiv nb213nf_izH1, 80.equiv nb213nf_ixH2, 96.equiv nb213nf_iyH2, 112.equiv nb213nf_izH2, 128.equiv nb213nf_ixM, 144.equiv nb213nf_iyM, 160.equiv nb213nf_izM, 176.equiv nb213nf_iqH, 192.equiv nb213nf_iqM, 208.equiv nb213nf_qqH, 224.equiv nb213nf_qqM, 240.equiv nb213nf_c6, 256.equiv nb213nf_c12, 272.equiv nb213nf_vctot, 288.equiv nb213nf_Vvdwtot, 304.equiv nb213nf_half, 320.equiv nb213nf_three, 336.equiv nb213nf_two, 352.equiv nb213nf_rinvH1, 368.equiv nb213nf_rinvH2, 384.equiv nb213nf_rinvM, 400.equiv nb213nf_krsqH1, 416.equiv nb213nf_krsqH2, 432.equiv nb213nf_krsqM, 448.equiv nb213nf_krf, 464.equiv nb213nf_crf, 480.equiv nb213nf_nri, 496.equiv nb213nf_iinr, 504.equiv nb213nf_jindex, 512.equiv nb213nf_jjnr, 520.equiv nb213nf_shift, 528.equiv nb213nf_shiftvec, 536.equiv nb213nf_facel, 544.equiv nb213nf_innerjjnr, 552.equiv nb213nf_is3, 560.equiv nb213nf_ii3, 564.equiv nb213nf_ntia, 568.equiv nb213nf_innerk, 572.equiv nb213nf_n, 576.equiv nb213nf_nn1, 580.equiv nb213nf_nouter, 584.equiv nb213nf_ninner, 588 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 600 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb213nf_nouter], eax mov [rsp + nb213nf_ninner], eax mov edi, [rdi] mov [rsp + nb213nf_nri], edi mov [rsp + nb213nf_iinr], rsi mov [rsp + nb213nf_jindex], rdx mov [rsp + nb213nf_jjnr], rcx mov [rsp + nb213nf_shift], r8 mov [rsp + nb213nf_shiftvec], r9 mov rsi, [rbp + nb213nf_p_facel] movsd xmm0, [rsi] movsd [rsp + nb213nf_facel], xmm0 mov rsi, [rbp + nb213nf_argkrf] mov rdi, [rbp + nb213nf_argcrf] movsd xmm1, [rsi] movsd xmm2, [rdi] shufpd xmm1, xmm1, 0 shufpd xmm2, xmm2, 0 movapd [rsp + nb213nf_krf], xmm1 movapd [rsp + nb213nf_crf], xmm2 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb213nf_half], eax mov [rsp + nb213nf_half+4], ebx movsd xmm1, [rsp + nb213nf_half] shufpd xmm1, xmm1, 0 ;# splat to all elements movapd xmm3, xmm1 addpd xmm3, xmm3 ;# one movapd xmm2, xmm3 addpd xmm2, xmm2 ;# two addpd xmm3, xmm2 ;# three movapd [rsp + nb213nf_half], xmm1 movapd [rsp + nb213nf_two], xmm2 movapd [rsp + nb213nf_three], xmm3 ;# assume we have at least one i particle - start directly mov rcx, [rsp + nb213nf_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx] ;# ebx =ii mov rdx, [rbp + nb213nf_charge] movsd xmm3, [rdx + rbx*8 + 8] movsd xmm4, [rdx + rbx*8 + 24] movsd xmm5, [rsp + nb213nf_facel] mulsd xmm3, xmm5 mulsd xmm4, xmm5 shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 movapd [rsp + nb213nf_iqH], xmm3 movapd [rsp + nb213nf_iqM], xmm4 mov rdx, [rbp + nb213nf_type] mov ecx, [rdx + rbx*4] shl ecx, 1 mov rdi, [rbp + nb213nf_p_ntype] imul ecx, [rdi] ;# rcx = ntia = 2*ntype*type[ii0] mov [rsp + nb213nf_ntia], ecx .nb213nf_threadloop: mov rsi, [rbp + nb213nf_count] ;# pointer to sync counter mov eax, [rsi].nb213nf_spinlock: mov ebx, eax ;# ebx=*count=nn0 add ebx, 1 ;# ebx=nn1=nn0+10 lock cmpxchg [esi], ebx ;# write nn1 to *counter, ;# if it hasnt changed. ;# or reread *counter to eax. pause ;# -> better p4 performance jnz .nb213nf_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb213nf_nri] mov edx, ecx sub ecx, ebx cmovle ebx, edx ;# if(nn1>nri) nn1=nri ;# Cleared the spinlock if we got here. ;# eax contains nn0, ebx contains nn1. mov [rsp + nb213nf_n], eax mov [rsp + nb213nf_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb213nf_outerstart jmp .nb213nf_end.nb213nf_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb213nf_nouter] mov [rsp + nb213nf_nouter], ebx.nb213nf_outer: mov rax, [rsp + nb213nf_shift] ;# rax = pointer into shift[] mov ebx, [rax+rsi*4] ;# rbx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb213nf_is3],ebx ;# store is3 mov rax, [rsp + nb213nf_shiftvec] ;# rax = base of shiftvec[] movsd xmm0, [rax + rbx*8] movsd xmm1, [rax + rbx*8 + 8] movsd xmm2, [rax + rbx*8 + 16] mov rcx, [rsp + nb213nf_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx+rsi*4] ;# ebx =ii movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 movapd xmm6, xmm0 movapd xmm7, xmm1 lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb213nf_pos] ;# rax = base of pos[] mov [rsp + nb213nf_ii3], ebx addsd xmm3, [rax + rbx*8] ;# ox addsd xmm4, [rax + rbx*8 + 8] ;# oy addsd xmm5, [rax + rbx*8 + 16] ;# oz addsd xmm6, [rax + rbx*8 + 24] ;# h1x addsd xmm7, [rax + rbx*8 + 32] ;# h1y shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 shufpd xmm6, xmm6, 0 shufpd xmm7, xmm7, 0 movapd [rsp + nb213nf_ixO], xmm3 movapd [rsp + nb213nf_iyO], xmm4 movapd [rsp + nb213nf_izO], xmm5 movapd [rsp + nb213nf_ixH1], xmm6 movapd [rsp + nb213nf_iyH1], xmm7 movsd xmm6, xmm2 movsd xmm3, xmm0 movsd xmm4, xmm1 movsd xmm5, xmm2 addsd xmm6, [rax + rbx*8 + 40] ;# h1z addsd xmm0, [rax + rbx*8 + 48] ;# h2x addsd xmm1, [rax + rbx*8 + 56] ;# h2y addsd xmm2, [rax + rbx*8 + 64] ;# h2z addsd xmm3, [rax + rbx*8 + 72] ;# mx addsd xmm4, [rax + rbx*8 + 80] ;# my addsd xmm5, [rax + rbx*8 + 88] ;# mz shufpd xmm6, xmm6, 0 shufpd xmm0, xmm0, 0 shufpd xmm1, xmm1, 0 shufpd xmm2, xmm2, 0 shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 movapd [rsp + nb213nf_izH1], xmm6 movapd [rsp + nb213nf_ixH2], xmm0 movapd [rsp + nb213nf_iyH2], xmm1 movapd [rsp + nb213nf_izH2], xmm2 movapd [rsp + nb213nf_ixM], xmm3 movapd [rsp + nb213nf_iyM], xmm4 movapd [rsp + nb213nf_izM], xmm5 ;# clear vctot xorpd xmm4, xmm4 movapd [rsp + nb213nf_vctot], xmm4 movapd [rsp + nb213nf_Vvdwtot], xmm4 mov rax, [rsp + nb213nf_jindex] mov ecx, [rax + rsi*4] ;# jindex[n] mov edx, [rax + rsi*4 + 4] ;# jindex[n+1] sub edx, ecx ;# number of innerloop atoms mov rsi, [rbp + nb213nf_pos] mov rdi, [rbp + nb213nf_faction] mov rax, [rsp + nb213nf_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb213nf_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 2 add ecx, [rsp + nb213nf_ninner] mov [rsp + nb213nf_ninner], ecx add edx, 0 mov [rsp + nb213nf_innerk], edx ;# number of innerloop atoms jge .nb213nf_unroll_loop jmp .nb213nf_checksingle.nb213nf_unroll_loop: ;# twice unrolled innerloop here mov rdx, [rsp + nb213nf_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] add qword ptr [rsp + nb213nf_innerjjnr], 8 ;# advance pointer (unrolled 2) mov rsi, [rbp + nb213nf_charge] ;# base of charge[] movlpd xmm3, [rsi + rax*8] movhpd xmm3, [rsi + rbx*8] movapd xmm4, xmm3 mulpd xmm3, [rsp + nb213nf_iqM] mulpd xmm4, [rsp + nb213nf_iqH] movd mm0, eax ;# use mmx registers as temp storage movd mm1, ebx movapd [rsp + nb213nf_qqM], xmm3 movapd [rsp + nb213nf_qqH], xmm4 mov rsi, [rbp + nb213nf_type] mov eax, [rsi + rax*4] mov ebx, [rsi + rbx*4] mov rsi, [rbp + nb213nf_vdwparam] shl eax, 1 ⌨️ 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?