nb_kernel230_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,465 行 · 第 1/3 页
S
1,465 行
movlpd xmm5, [rdi + rax*8 + 16] movhpd xmm3, [rdi + rbx*8] movhpd xmm4, [rdi + rbx*8 + 8] movhpd xmm5, [rdi + rbx*8 + 16] addpd xmm3, xmm9 addpd xmm4, xmm10 addpd xmm5, xmm11 movlpd [rdi + rax*8], xmm3 movlpd [rdi + rax*8 + 8], xmm4 movlpd [rdi + rax*8 + 16], xmm5 movhpd [rdi + rbx*8], xmm3 movhpd [rdi + rbx*8 + 8], xmm4 movhpd [rdi + rbx*8 + 16], xmm5 ;# should we do one more iteration? sub dword ptr [rsp + nb230_innerk], 2 jl .nb230_checksingle jmp .nb230_unroll_loop.nb230_checksingle: mov edx, [rsp + nb230_innerk] and edx, 1 jnz .nb230_dosingle jmp .nb230_updateouterdata.nb230_dosingle: mov rsi, [rbp + nb230_charge] mov rdi, [rbp + nb230_pos] mov rcx, [rsp + nb230_innerjjnr] mov eax, [rcx] mov rsi, [rbp + nb230_charge] ;# base of charge[] movsd xmm3, [rsi + rax*8] mulsd xmm3, [rsp + nb230_iq] ;# qq mov rsi, [rbp + nb230_type] mov r8d, [rsi + rax*4] mov rsi, [rbp + nb230_vdwparam] shl r8d, 1 mov edi, [rsp + nb230_ntia] add r8d, edi movsd xmm4, [rsi + r8*8] movsd xmm6, [rsi + r8*8 + 8] movapd [rsp + nb230_c6], xmm4 movapd [rsp + nb230_c12], xmm6 mov rsi, [rbp + nb230_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 ;# move two coordinates to xmm4-xmm6 movsd xmm4, [rsi + rax*8] movsd xmm5, [rsi + rax*8 + 8] movsd xmm6, [rsi + rax*8 + 16] ;# calc dr subsd xmm4, [rsp + nb230_ix] subsd xmm5, [rsp + nb230_iy] subsd xmm6, [rsp + nb230_iz] ;# store dr movapd [rsp + nb230_dx], xmm4 movapd [rsp + nb230_dy], xmm5 movapd [rsp + nb230_dz], xmm6 ;# square it mulsd xmm4,xmm4 mulsd xmm5,xmm5 mulsd xmm6,xmm6 addsd xmm4, xmm5 addsd xmm4, xmm6 ;# rsq in xmm4 cvtsd2ss xmm5, xmm4 rsqrtss xmm5, xmm5 cvtss2sd xmm2, xmm5 ;# lu in low xmm2 movapd xmm7, [rsp + nb230_krf] ;# lookup seed in xmm2 movapd xmm5, xmm2 ;# copy of lu mulsd xmm2, xmm2 ;# lu*lu movapd xmm1, [rsp + nb230_three] mulsd xmm7, xmm4 ;# krsq mulsd xmm2, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb230_half] subsd xmm1, xmm2 ;# 30-rsq*lu*lu mulsd xmm1, xmm5 mulsd xmm1, xmm0 ;# xmm0=iter1 of rinv (new lu) movapd xmm5, xmm1 ;# copy of lu mulsd xmm1, xmm1 ;# lu*lu movapd xmm2, [rsp + nb230_three] mulsd xmm1, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb230_half] subsd xmm2, xmm1 ;# 30-rsq*lu*lu mulsd xmm2, xmm5 mulsd xmm0, xmm2 ;# xmm0=rinv movapd xmm6, xmm0 addsd xmm6, xmm7 ;# xmm6=rinv+ krsq movapd xmm1, xmm4 subsd xmm6, [rsp + nb230_crf] mulsd xmm6, xmm3 mulsd xmm7, [rsp + nb230_two] movapd xmm1, xmm0 subsd xmm1, xmm7 ;# rinv-2*krsq mulsd xmm1, xmm0 ;# (rinv-2*krsq)*rinv mulsd xmm3, xmm1 ;# qq*(rinv-2*krsq)*rinv ;# xmm3=fstmp addsd xmm6, [rsp + nb230_vctot] movsd [rsp + nb230_vctot], xmm6 ;# LJ table interaction. xmm0=rinv, xmm4=rsq mulsd xmm4, xmm0 ;# xmm4=r mulsd xmm4, [rsp + nb230_tsc] cvttsd2si r8d, xmm4 ;# mm6 = lu idx cvtsi2sd xmm5, r8d subsd xmm4, xmm5 movapd xmm1, xmm4 ;# xmm1=eps movapd xmm2, xmm1 mulsd xmm2, xmm2 ;# xmm2=eps2 shl r8d, 3 ;# idx *= 8 mov rsi, [rbp + nb230_VFtab] ;# load both disp. and rep. tables in parallel movsd xmm4, [rsi + r8*8] movsd xmm5, [rsi + r8*8 + 8] movsd xmm6, [rsi + r8*8 + 16] movsd xmm7, [rsi + r8*8 + 24] movsd xmm8, [rsi + r8*8 + 32] movsd xmm9, [rsi + r8*8 + 40] movsd xmm10, [rsi + r8*8 + 48] movsd xmm11, [rsi + r8*8 + 56] ;# dispersion table ready in xmm4-xmm7, repulsion in xmm8-xmm11 mulsd xmm7, xmm1 ;# Heps mulsd xmm11, xmm1 mulsd xmm6, xmm1 ;# Geps mulsd xmm10, xmm1 mulsd xmm7, xmm1 ;# Heps2 mulsd xmm11, xmm1 addsd xmm5, xmm6 ;# F+Geps addsd xmm9, xmm10 addsd xmm5, xmm7 ;# F+Geps+Heps2 = Fp addsd xmm9, xmm11 addsd xmm7, xmm7 ;# 2*Heps2 addsd xmm11, xmm11 addsd xmm7, xmm6 ;# 2*Heps2+Geps addsd xmm11, xmm10 addsd xmm7, xmm5 ;# FF = Fp + 2*Heps2 + Geps addsd xmm11, xmm9 mulsd xmm5, xmm1 ;# eps*Fp mulsd xmm9, xmm1 movapd xmm12, [rsp + nb230_c6] movapd xmm13, [rsp + nb230_c12] addsd xmm5, xmm4 ;# VV addsd xmm9, xmm8 mulsd xmm5, xmm12 ;# VV*c6 = vnb6 mulsd xmm9, xmm13 ;# VV*c12 = vnb12 addsd xmm5, xmm9 addsd xmm5, [rsp + nb230_Vvdwtot] movsd [rsp + nb230_Vvdwtot], xmm5 mulsd xmm7, xmm12 ;# FF*c6 = fnb6 mulsd xmm11, xmm13 ;# FF*c12 = fnb12 addsd xmm7, xmm11 mulsd xmm7, [rsp + nb230_tsc] subsd xmm3, xmm7 mulsd xmm3, xmm0 ;# fscal movapd xmm9, xmm3 movapd xmm10, xmm3 movapd xmm11, xmm3 movapd xmm12, [rsp + nb230_fix] movapd xmm13, [rsp + nb230_fiy] movapd xmm14, [rsp + nb230_fiz] mulsd xmm9, [rsp + nb230_dx] mulsd xmm10, [rsp + nb230_dy] mulsd xmm11, [rsp + nb230_dz] ;# accumulate i forces addsd xmm12, xmm9 addsd xmm13, xmm10 addsd xmm14, xmm11 movsd [rsp + nb230_fix], xmm12 movsd [rsp + nb230_fiy], xmm13 movsd [rsp + nb230_fiz], xmm14 ;# the fj's - start by accumulating forces from memory mov rdi, [rbp + nb230_faction] addsd xmm9, [rdi + rax*8] addsd xmm10, [rdi + rax*8 + 8] addsd xmm11, [rdi + rax*8 + 16] movsd [rdi + rax*8], xmm9 movsd [rdi + rax*8 + 8], xmm10 movsd [rdi + rax*8 + 16], xmm11 .nb230_updateouterdata: mov ecx, [rsp + nb230_ii3] mov rdi, [rbp + nb230_faction] mov rsi, [rbp + nb230_fshift] mov edx, [rsp + nb230_is3] ;# accumulate i forces in xmm0, xmm1, xmm2 movapd xmm0, [rsp + nb230_fix] movapd xmm1, [rsp + nb230_fiy] movapd xmm2, [rsp + nb230_fiz] movhlps xmm3, xmm0 movhlps xmm4, xmm1 movhlps xmm5, xmm2 addpd xmm0, xmm3 addpd xmm1, xmm4 addpd 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 ;# increment fshift force movsd xmm3, [rsi + rdx*8] movsd xmm4, [rsi + rdx*8 + 8] movsd xmm5, [rsi + rdx*8 + 16] subsd xmm3, xmm0 subsd xmm4, xmm1 subsd xmm5, xmm2 movsd [rsi + rdx*8], xmm3 movsd [rsi + rdx*8 + 8], xmm4 movsd [rsi + rdx*8 + 16], xmm5 ;# get n from stack mov esi, [rsp + nb230_n] ;# get group index for i particle mov rdx, [rbp + nb230_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movapd xmm7, [rsp + nb230_vctot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# low xmm7 has the sum now ;# add earlier value from mem mov rax, [rbp + nb230_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 + nb230_Vvdwtot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# low xmm7 has the sum now ;# add earlier value from mem mov rax, [rbp + nb230_Vvdw] addsd xmm7, [rax + rdx*8] ;# move back to mem movsd [rax + rdx*8], xmm7 ;# finish if last mov ecx, [rsp + nb230_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb230_outerend ;# not last, iterate outer loop once more! mov [rsp + nb230_n], esi jmp .nb230_outer.nb230_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb230_nri] ;# esi already loaded with n above sub ecx, esi jz .nb230_end ;# non-zero, do one more workunit jmp .nb230_threadloop.nb230_end: mov eax, [rsp + nb230_nouter] mov ebx, [rsp + nb230_ninner] mov rcx, [rbp + nb230_outeriter] mov rdx, [rbp + nb230_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 456 emms pop r15 pop r14 pop r13 pop r12 pop rbx pop rbp ret .globl nb_kernel230nf_x86_64_sse2.globl _nb_kernel230nf_x86_64_sse2nb_kernel230nf_x86_64_sse2: _nb_kernel230nf_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb230nf_fshift, 16.equiv nb230nf_gid, 24.equiv nb230nf_pos, 32.equiv nb230nf_faction, 40.equiv nb230nf_charge, 48.equiv nb230nf_p_facel, 56.equiv nb230nf_argkrf, 64.equiv nb230nf_argcrf, 72.equiv nb230nf_Vc, 80.equiv nb230nf_type, 88.equiv nb230nf_p_ntype, 96.equiv nb230nf_vdwparam, 104.equiv nb230nf_Vvdw, 112.equiv nb230nf_p_tabscale, 120.equiv nb230nf_VFtab, 128.equiv nb230nf_invsqrta, 136.equiv nb230nf_dvda, 144.equiv nb230nf_p_gbtabscale, 152.equiv nb230nf_GBtab, 160.equiv nb230nf_p_nthreads, 168.equiv nb230nf_count, 176.equiv nb230nf_mtx, 184.equiv nb230nf_outeriter, 192.equiv nb230nf_inneriter, 200.equiv nb230nf_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb230nf_ix, 0.equiv nb230nf_iy, 16.equiv nb230nf_iz, 32.equiv nb230nf_iq, 48.equiv nb230nf_c6, 64.equiv nb230nf_c12, 80.equiv nb230nf_vctot, 96.equiv nb230nf_Vvdwtot, 112.equiv nb230nf_half, 128.equiv nb230nf_three, 144.equiv nb230nf_krf, 160.equiv nb230nf_crf, 176.equiv nb230nf_tsc, 192.equiv nb230nf_nri, 208.equiv nb230nf_iinr, 216.equiv nb230nf_jindex, 224.equiv nb230nf_jjnr, 232.equiv nb230nf_shift, 240.equiv nb230nf_shiftvec, 248.equiv nb230nf_facel, 256.equiv nb230nf_innerjjnr, 264.equiv nb230nf_is3, 272.equiv nb230nf_ii3, 280.equiv nb230nf_ntia, 284.equiv nb230nf_innerk, 288.equiv nb230nf_n, 292.equiv nb230nf_nn1, 296.equiv nb230nf_ntype, 300.equiv nb230nf_nouter, 304.equiv nb230nf_ninner, 308 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 328 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb230nf_nouter], eax mov [rsp + nb230nf_ninner], eax mov edi, [rdi] mov [rsp + nb230nf_nri], edi mov [rsp + nb230nf_iinr], rsi mov [rsp + nb230nf_jindex], rdx mov [rsp + nb230nf_jjnr], rcx mov [rsp + nb230nf_shift], r8 mov [rsp + nb230nf_shiftvec], r9 mov rdi, [rbp + nb230nf_p_ntype] mov edi, [rdi] mov [rsp + nb230nf_ntype], edi mov rsi, [rbp + nb230nf_p_facel] movsd xmm0, [rsi] movsd [rsp + nb230nf_facel], xmm0 mov rax, [rbp + nb230nf_p_tabscale] movsd xmm3, [rax] shufpd xmm3, xmm3, 0 movapd [rsp + nb230nf_tsc], xmm3 mov rsi, [rbp + nb230nf_argkrf] mov rdi, [rbp + nb230nf_argcrf] movsd xmm1, [rsi] movsd xmm2, [rdi] shufpd xmm1, xmm1, 0 shufpd xmm2, xmm2, 0 movapd [rsp + nb230nf_krf], xmm1 movapd [rsp + nb230nf_crf], xmm2 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb230nf_half], eax mov [rsp + nb230nf_half+4], ebx movsd xmm1, [rsp + nb230nf_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 + nb230nf_half], xmm1 movapd [rsp + nb230nf_three], xmm3.nb230nf_threadloop: mov rsi, [rbp + nb230nf_count] ;# pointer to sync counter mov eax, [rsi].nb230nf_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 .nb230nf_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb230nf_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 + nb230nf_n], eax mov [rsp + nb230nf_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb230nf_outerstart jmp .nb230nf_end.nb230nf_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb230nf_nouter] mov [rsp + nb230nf_nouter], ebx.nb230nf_outer: mov rax, [rsp + nb230nf_shift] ;# eax = pointer into shift[] mov ebx, [rax+rsi*4] ;# ebx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb230nf_is3],ebx ;# store is3 mov rax, [rsp + nb230nf_shiftvec] ;# eax = base of shiftvec[] movsd xmm0, [rax + rbx*8] movsd xmm1, [rax + rbx*8 + 8] movsd xmm2, [rax + rbx*8 + 16] ⌨️ 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?