nb_kernel101_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,627 行 · 第 1/3 页
S
1,627 行
movlpd [rdi + rax*8], xmm0 movlpd [rdi + rax*8 + 8], xmm1 movlpd [rdi + rax*8 + 16], xmm2 movhpd [rdi + rbx*8], xmm0 movhpd [rdi + rbx*8 + 8], xmm1 movhpd [rdi + rbx*8 + 16], xmm2 ;# should we do one more iteration? sub dword ptr [rsp + nb101_innerk], 2 jl .nb101_checksingle jmp .nb101_unroll_loop.nb101_checksingle: mov edx, [rsp + nb101_innerk] and edx, 1 jnz .nb101_dosingle jmp .nb101_updateouterdata.nb101_dosingle: mov rdx, [rsp + nb101_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov rsi, [rbp + nb101_charge] ;# base of charge[] xorpd xmm6, xmm6 movlpd xmm6, [rsi + rax*8] ;# jq A movapd xmm3, [rsp + nb101_iqO] movapd xmm4, [rsp + nb101_iqH] mulsd xmm3, xmm6 ;# qqO mulsd xmm4, xmm6 ;# qqH movapd [rsp + nb101_qqO], xmm3 movapd [rsp + nb101_qqH], xmm4 mov rsi, [rbp + nb101_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 ;# move coordinates to xmm0-xmm2 movlpd xmm4, [rsi + rax*8] movlpd xmm5, [rsi + rax*8 + 8] movlpd xmm6, [rsi + rax*8 + 16] movapd xmm0, xmm4 movapd xmm1, xmm5 movapd xmm2, xmm6 ;# calc dr subsd xmm4, [rsp + nb101_ixO] subsd xmm5, [rsp + nb101_iyO] subsd xmm6, [rsp + nb101_izO] ;# store dr movapd [rsp + nb101_dxO], xmm4 movapd [rsp + nb101_dyO], xmm5 movapd [rsp + nb101_dzO], xmm6 ;# square it mulsd xmm4,xmm4 mulsd xmm5,xmm5 mulsd xmm6,xmm6 addsd xmm4, xmm5 addsd xmm4, xmm6 movapd xmm7, xmm4 ;# rsqO in xmm7 ;# move j coords to xmm4-xmm6 movapd xmm4, xmm0 movapd xmm5, xmm1 movapd xmm6, xmm2 ;# calc dr subsd xmm4, [rsp + nb101_ixH1] subsd xmm5, [rsp + nb101_iyH1] subsd xmm6, [rsp + nb101_izH1] ;# store dr movapd [rsp + nb101_dxH1], xmm4 movapd [rsp + nb101_dyH1], xmm5 movapd [rsp + nb101_dzH1], xmm6 ;# square it mulsd xmm4,xmm4 mulsd xmm5,xmm5 mulsd xmm6,xmm6 addsd xmm6, xmm5 addsd xmm6, xmm4 ;# rsqH1 in xmm6 ;# move j coords to xmm3-xmm5 movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 ;# calc dr subsd xmm3, [rsp + nb101_ixH2] subsd xmm4, [rsp + nb101_iyH2] subsd xmm5, [rsp + nb101_izH2] ;# store dr movapd [rsp + nb101_dxH2], xmm3 movapd [rsp + nb101_dyH2], xmm4 movapd [rsp + nb101_dzH2], xmm5 ;# square it mulsd xmm3,xmm3 mulsd xmm4,xmm4 mulsd xmm5,xmm5 addsd xmm5, xmm4 addsd xmm5, xmm3 ;# rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 ;# start with rsqO - put seed in xmm2 cvtsd2ss xmm2, xmm7 rsqrtss xmm2, xmm2 cvtss2sd xmm2, xmm2 movapd xmm3, xmm2 mulsd xmm2, xmm2 movapd xmm4, [rsp + nb101_three] mulsd xmm2, xmm7 ;# rsq*lu*lu subsd xmm4, xmm2 ;# 30-rsq*lu*lu mulsd xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulsd xmm4, [rsp + nb101_half] ;# iter1 ( new lu) movapd xmm3, xmm4 mulsd xmm4, xmm4 ;# lu*lu mulsd xmm7, xmm4 ;# rsq*lu*lu movapd xmm4, [rsp + nb101_three] subsd xmm4, xmm7 ;# 3-rsq*lu*lu mulsd xmm4, xmm3 ;# lu*( 3-rsq*lu*lu) mulsd xmm4, [rsp + nb101_half] ;# rinv movapd xmm7, xmm4 ;# rinvO in xmm7 ;# rsqH1 - seed in xmm2 cvtsd2ss xmm2, xmm6 rsqrtss xmm2, xmm2 cvtss2sd xmm2, xmm2 movapd xmm3, xmm2 mulsd xmm2, xmm2 movapd xmm4, [rsp + nb101_three] mulsd xmm2, xmm6 ;# rsq*lu*lu subsd xmm4, xmm2 ;# 30-rsq*lu*lu mulsd xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulsd xmm4, [rsp + nb101_half] ;# iter1 ( new lu) movapd xmm3, xmm4 mulsd xmm4, xmm4 ;# lu*lu mulsd xmm6, xmm4 ;# rsq*lu*lu movapd xmm4, [rsp + nb101_three] subsd xmm4, xmm6 ;# 3-rsq*lu*lu mulsd xmm4, xmm3 ;# lu*( 3-rsq*lu*lu) mulsd xmm4, [rsp + nb101_half] ;# rinv movapd xmm6, xmm4 ;# rinvH1 in xmm6 ;# rsqH2 - seed in xmm2 cvtsd2ss xmm2, xmm5 rsqrtss xmm2, xmm2 cvtss2sd xmm2, xmm2 movapd xmm3, xmm2 mulsd xmm2, xmm2 movapd xmm4, [rsp + nb101_three] mulsd xmm2, xmm5 ;# rsq*lu*lu subsd xmm4, xmm2 ;# 30-rsq*lu*lu mulsd xmm4, xmm3 ;# lu*(3-rsq*lu*lu) mulsd xmm4, [rsp + nb101_half] ;# iter1 ( new lu) movapd xmm3, xmm4 mulsd xmm4, xmm4 ;# lu*lu mulsd xmm5, xmm4 ;# rsq*lu*lu movapd xmm4, [rsp + nb101_three] subsd xmm4, xmm5 ;# 3-rsq*lu*lu mulsd xmm4, xmm3 ;# lu*( 3-rsq*lu*lu) mulsd xmm4, [rsp + nb101_half] ;# rinv movapd xmm5, xmm4 ;# rinvH2 in xmm5 ;# do O interactions movapd xmm4, xmm7 mulsd xmm4, xmm4 ;# xmm7=rinv, xmm4=rinvsq mulsd xmm7, [rsp + nb101_qqO] ;# xmm7=vcoul mulsd xmm4, xmm7 ;# total fsO in xmm4 addsd xmm7, [rsp + nb101_vctot] movlpd [rsp + nb101_vctot], xmm7 movapd xmm0, [rsp + nb101_dxO] movapd xmm1, [rsp + nb101_dyO] movapd xmm2, [rsp + nb101_dzO] mulsd xmm0, xmm4 mulsd xmm1, xmm4 mulsd xmm2, xmm4 ;# update O forces movapd xmm3, [rsp + nb101_fixO] movapd xmm4, [rsp + nb101_fiyO] movapd xmm7, [rsp + nb101_fizO] addsd xmm3, xmm0 addsd xmm4, xmm1 addsd xmm7, xmm2 movlpd [rsp + nb101_fixO], xmm3 movlpd [rsp + nb101_fiyO], xmm4 movlpd [rsp + nb101_fizO], xmm7 ;# update j forces with water O movlpd [rsp + nb101_fjx], xmm0 movlpd [rsp + nb101_fjy], xmm1 movlpd [rsp + nb101_fjz], xmm2 ;# H1 interactions movapd xmm4, xmm6 mulsd xmm4, xmm4 ;# xmm6=rinv, xmm4=rinvsq mulsd xmm6, [rsp + nb101_qqH] ;# xmm6=vcoul mulsd xmm4, xmm6 ;# total fsH1 in xmm4 addsd xmm6, [rsp + nb101_vctot] movapd xmm0, [rsp + nb101_dxH1] movapd xmm1, [rsp + nb101_dyH1] movapd xmm2, [rsp + nb101_dzH1] movlpd [rsp + nb101_vctot], xmm6 mulsd xmm0, xmm4 mulsd xmm1, xmm4 mulsd xmm2, xmm4 ;# update H1 forces movapd xmm3, [rsp + nb101_fixH1] movapd xmm4, [rsp + nb101_fiyH1] movapd xmm7, [rsp + nb101_fizH1] addsd xmm3, xmm0 addsd xmm4, xmm1 addsd xmm7, xmm2 movlpd [rsp + nb101_fixH1], xmm3 movlpd [rsp + nb101_fiyH1], xmm4 movlpd [rsp + nb101_fizH1], xmm7 ;# update j forces with water H1 addsd xmm0, [rsp + nb101_fjx] addsd xmm1, [rsp + nb101_fjy] addsd xmm2, [rsp + nb101_fjz] movsd [rsp + nb101_fjx], xmm0 movsd [rsp + nb101_fjy], xmm1 movsd [rsp + nb101_fjz], xmm2 ;# H2 interactions movapd xmm4, xmm5 mulsd xmm4, xmm4 ;# xmm5=rinv, xmm4=rinvsq mulsd xmm5, [rsp + nb101_qqH] ;# xmm5=vcoul mulsd xmm4, xmm5 ;# total fsH1 in xmm4 addsd xmm5, [rsp + nb101_vctot] movapd xmm0, [rsp + nb101_dxH2] movapd xmm1, [rsp + nb101_dyH2] movapd xmm2, [rsp + nb101_dzH2] movlpd [rsp + nb101_vctot], xmm5 mulsd xmm0, xmm4 mulsd xmm1, xmm4 mulsd xmm2, xmm4 ;# update H2 forces movapd xmm3, [rsp + nb101_fixH2] movapd xmm4, [rsp + nb101_fiyH2] movapd xmm7, [rsp + nb101_fizH2] addsd xmm3, xmm0 addsd xmm4, xmm1 addsd xmm7, xmm2 movlpd [rsp + nb101_fixH2], xmm3 movlpd [rsp + nb101_fiyH2], xmm4 movlpd [rsp + nb101_fizH2], xmm7 mov rdi, [rbp + nb101_faction] ;# update j forces addsd xmm0, [rsp + nb101_fjx] addsd xmm1, [rsp + nb101_fjy] addsd xmm2, [rsp + nb101_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.nb101_updateouterdata: mov ecx, [rsp + nb101_ii3] mov rdi, [rbp + nb101_faction] mov rsi, [rbp + nb101_fshift] mov edx, [rsp + nb101_is3] ;# accumulate Oi forces in xmm0, xmm1, xmm2 movapd xmm0, [rsp + nb101_fixO] movapd xmm1, [rsp + nb101_fiyO] movapd xmm2, [rsp + nb101_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 + nb101_fixH1] movapd xmm1, [rsp + nb101_fiyH1] movapd xmm2, [rsp + nb101_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 + nb101_fixH2] movapd xmm1, [rsp + nb101_fiyH2] movapd xmm2, [rsp + nb101_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 movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, 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 ;# 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 + nb101_n] ;# get group index for i particle mov rdx, [rbp + nb101_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movapd xmm7, [rsp + nb101_vctot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# pos 0-1 in xmm7 have the sum now ;# add earlier value from mem mov rax, [rbp + nb101_Vc] addsd xmm7, [rax + rdx*8] ;# move back to mem movsd [rax + rdx*8], xmm7 ;# finish if last mov ecx, [rsp + nb101_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb101_outerend ;# not last, iterate outer loop once more! mov [rsp + nb101_n], esi jmp .nb101_outer.nb101_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb101_nri] ;# esi already loaded with n above sub ecx, esi jz .nb101_end ;# non-zero, do one more workunit jmp .nb101_threadloop.nb101_end: emms mov eax, [rsp + nb101_nouter] mov ebx, [rsp + nb101_ninner] mov rcx, [rbp + nb101_outeriter] mov rdx, [rbp + nb101_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 696 pop r15 pop r14 pop r13 pop r12 pop rbx pop rbp ret.globl nb_kernel101nf_x86_64_sse2.globl _nb_kernel101nf_x86_64_sse2nb_kernel101nf_x86_64_sse2: _nb_kernel101nf_x86_64_sse2: .equiv nb101nf_fshift, 16.equiv nb101nf_gid, 24.equiv nb101nf_pos, 32.equiv nb101nf_faction, 40.equiv nb101nf_charge, 48.equiv nb101nf_p_facel, 56.equiv nb101nf_argkrf, 64.equiv nb101nf_argcrf, 72.equiv nb101nf_Vc, 80.equiv nb101nf_type, 88.equiv nb101nf_p_ntype, 96.equiv nb101nf_vdwparam, 104.equiv nb101nf_Vvdw, 112.equiv nb101nf_p_tabscale, 120.equiv nb101nf_VFtab, 128.equiv nb101nf_invsqrta, 136.equiv nb101nf_dvda, 144.equiv nb101nf_p_gbtabscale, 152.equiv nb101nf_GBtab, 160.equiv nb101nf_p_nthreads, 168.equiv nb101nf_count, 176.equiv nb101nf_mtx, 184.equiv nb101nf_outeriter, 192.equiv nb101nf_inneriter, 200.equiv nb101nf_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb101nf_ixO, 0.equiv nb101nf_iyO, 16.equiv nb101nf_izO, 32.equiv nb101nf_ixH1, 48.equiv nb101nf_iyH1, 64.equiv nb101nf_izH1, 80.equiv nb101nf_ixH2, 96.equiv nb101nf_iyH2, 112.equiv nb101nf_izH2, 128.equiv nb101nf_iqO, 144.equiv nb101nf_iqH, 160.equiv nb101nf_qqO, 176.equiv nb101nf_qqH, 192.equiv nb101nf_vctot, 208.equiv nb101nf_half, 224.equiv nb101nf_three, 240.equiv nb101nf_is3, 256.equiv nb101nf_ii3, 260.equiv nb101nf_nri, 264.equiv nb101nf_iinr, 272.equiv nb101nf_jindex, 280.equiv nb101nf_jjnr, 288.equiv nb101nf_shift, 296.equiv nb101nf_shiftvec, 304.equiv nb101nf_facel, 312.equiv nb101nf_innerjjnr, 320.equiv nb101nf_innerk, 328.equiv nb101nf_n, 332.equiv nb101nf_nn1, 336.equiv nb101nf_nouter, 340.equiv nb101nf_ninner, 344 push rbp mov rbp, rsp push rbx emms ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb101nf_nouter], eax mov [rsp + nb101nf_ninner], eax push r12 push r13 push r14 push r15 sub rsp, 360 ;# local variable stack space (n*16+8) mov edi, [rdi] mov [rsp + nb101nf_nri], edi mov [rsp + nb101nf_iinr], rsi mov [rsp + nb101nf_jindex], rdx mov [rsp + nb101nf_jjnr], rcx mov [rsp + nb101nf_shift], r8 mov [rsp + nb101nf_shiftvec], r9 mov rsi, [rbp + nb101nf_p_facel] movsd xmm0, [rsi] movsd [rsp + nb101nf_facel], 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?