nb_kernel030_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,306 行 · 第 1/3 页
S
1,306 行
movhpd xmm5, [rdi + rbx*8 + 16] movapd xmm12, [rsp + nb030_fix] movapd xmm13, [rsp + nb030_fiy] movapd xmm14, [rsp + nb030_fiz] mulpd xmm9, [rsp + nb030_dx] mulpd xmm10, [rsp + nb030_dy] mulpd xmm11, [rsp + nb030_dz] ;# accumulate i forces addpd xmm12, xmm9 addpd xmm13, xmm10 addpd xmm14, xmm11 movapd [rsp + nb030_fix], xmm12 movapd [rsp + nb030_fiy], xmm13 movapd [rsp + nb030_fiz], xmm14 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 + nb030_innerk], 2 jl .nb030_checksingle jmp .nb030_unroll_loop.nb030_checksingle: mov edx, [rsp + nb030_innerk] and edx, 1 jnz .nb030_dosingle jmp .nb030_updateouterdata.nb030_dosingle: mov rdx, [rsp + nb030_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov rsi, [rbp + nb030_type] mov r8d, [rsi + rax*4] mov rsi, [rbp + nb030_vdwparam] shl r8d, 1 mov edi, [rsp + nb030_ntia] add r8d, edi movsd xmm4, [rsi + r8*8] movsd xmm6, [rsi + r8*8 + 8] movapd [rsp + nb030_c6], xmm4 movapd [rsp + nb030_c12], xmm6 mov rsi, [rbp + nb030_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 + nb030_ix] subsd xmm5, [rsp + nb030_iy] subsd xmm6, [rsp + nb030_iz] ;# store dr movapd [rsp + nb030_dx], xmm4 movapd [rsp + nb030_dy], xmm5 movapd [rsp + nb030_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 ;# lookup seed in xmm2 movapd xmm5, xmm2 ;# copy of lu mulsd xmm2, xmm2 ;# lu*lu movapd xmm1, [rsp + nb030_three] mulsd xmm2, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb030_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 + nb030_three] mulsd xmm1, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb030_half] subsd xmm2, xmm1 ;# 30-rsq*lu*lu mulsd xmm2, xmm5 mulsd xmm2, xmm0 ;# xmm0=iter2 of rinv (new lu) mulsd xmm4, xmm2 ;# xmm4=r mulsd xmm4, [rsp + nb030_tsc] cvttsd2si r10d, xmm4 ;# mm6 = lu idx cvtsi2sd xmm5, r10d subsd xmm4, xmm5 movapd xmm1, xmm4 ;# xmm1=eps ;# xmm2=rinv movapd xmm3, xmm4 ;# eps shl r10d, 3 ;# idx *= 8 mov rsi, [rbp + nb030_VFtab] ;# indices in r10, r11. Load dispersion and repulsion tables in parallel. movsd xmm4, [rsi + r10*8] movsd xmm5, [rsi + r10*8 + 8] movsd xmm6, [rsi + r10*8 + 16] movsd xmm7, [rsi + r10*8 + 24] movsd xmm8, [rsi + r10*8 + 32] movsd xmm9, [rsi + r10*8 + 40] movsd xmm10, [rsi + r10*8 + 48] movsd xmm11, [rsi + r10*8 + 56] ;# tables ready, in xmm4-xmm7 and 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 + nb030_c6] movapd xmm13, [rsp + nb030_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 + nb030_Vvdwtot] movsd [rsp + nb030_Vvdwtot], xmm5 mulsd xmm7, xmm12 ;# FF*c6 = fnb6 mulsd xmm11, xmm13 ;# FF*c12 = fnb12 addsd xmm7, xmm11 mulsd xmm7, [rsp + nb030_tsc] mulsd xmm7, xmm2 xorpd xmm9, xmm9 subsd xmm9, xmm7 movapd xmm10, xmm9 movapd xmm11, xmm9 movapd xmm12, [rsp + nb030_fix] movapd xmm13, [rsp + nb030_fiy] movapd xmm14, [rsp + nb030_fiz] mulsd xmm9, [rsp + nb030_dx] mulsd xmm10, [rsp + nb030_dy] mulsd xmm11, [rsp + nb030_dz] ;# accumulate i forces addsd xmm12, xmm9 addsd xmm13, xmm10 addsd xmm14, xmm11 movsd [rsp + nb030_fix], xmm12 movsd [rsp + nb030_fiy], xmm13 movsd [rsp + nb030_fiz], xmm14 ;# the fj's - start by accumulating forces from memory mov rdi, [rbp + nb030_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 .nb030_updateouterdata: mov ecx, [rsp + nb030_ii3] mov rdi, [rbp + nb030_faction] mov rsi, [rbp + nb030_fshift] mov edx, [rsp + nb030_is3] ;# accumulate i forces in xmm0, xmm1, xmm2 movapd xmm0, [rsp + nb030_fix] movapd xmm1, [rsp + nb030_fiy] movapd xmm2, [rsp + nb030_fiz] 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 ;# 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 + nb030_n] ;# get group index for i particle mov rdx, [rbp + nb030_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total lj energy and update it movapd xmm7, [rsp + nb030_Vvdwtot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# low xmm7 has the sum now ;# add earlier value from mem mov rax, [rbp + nb030_Vvdw] addsd xmm7, [rax + rdx*8] ;# move back to mem movsd [rax + rdx*8], xmm7 ;# finish if last mov ecx, [rsp + nb030_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb030_outerend ;# not last, iterate outer loop once more! mov [rsp + nb030_n], esi jmp .nb030_outer.nb030_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb030_nri] ;# esi already loaded with n above sub ecx, esi jz .nb030_end ;# non-zero, do one more workunit jmp .nb030_threadloop.nb030_end: mov eax, [rsp + nb030_nouter] mov ebx, [rsp + nb030_ninner] mov rcx, [rbp + nb030_outeriter] mov rdx, [rbp + nb030_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 376 emms pop r15 pop r14 pop r13 pop r12 pop rbx pop rbp ret.globl nb_kernel030nf_x86_64_sse2.globl _nb_kernel030nf_x86_64_sse2nb_kernel030nf_x86_64_sse2: _nb_kernel030nf_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb030nf_fshift, 16.equiv nb030nf_gid, 24.equiv nb030nf_pos, 32.equiv nb030nf_faction, 40.equiv nb030nf_charge, 48.equiv nb030nf_p_facel, 56.equiv nb030nf_argkrf, 64.equiv nb030nf_argcrf, 72.equiv nb030nf_Vc, 80.equiv nb030nf_type, 88.equiv nb030nf_p_ntype, 96.equiv nb030nf_vdwparam, 104.equiv nb030nf_Vvdw, 112.equiv nb030nf_p_tabscale, 120.equiv nb030nf_VFtab, 128.equiv nb030nf_invsqrta, 136.equiv nb030nf_dvda, 144.equiv nb030nf_p_gbtabscale, 152.equiv nb030nf_GBtab, 160.equiv nb030nf_p_nthreads, 168.equiv nb030nf_count, 176.equiv nb030nf_mtx, 184.equiv nb030nf_outeriter, 192.equiv nb030nf_inneriter, 200.equiv nb030nf_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb030nf_ix, 0.equiv nb030nf_iy, 16.equiv nb030nf_iz, 32.equiv nb030nf_tsc, 48.equiv nb030nf_c6, 64.equiv nb030nf_c12, 80.equiv nb030nf_Vvdwtot, 96.equiv nb030nf_half, 112.equiv nb030nf_three, 128.equiv nb030nf_is3, 144.equiv nb030nf_ii3, 148.equiv nb030nf_nri, 152.equiv nb030nf_iinr, 160.equiv nb030nf_jindex, 168.equiv nb030nf_jjnr, 176.equiv nb030nf_shift, 184.equiv nb030nf_shiftvec, 192.equiv nb030nf_innerjjnr, 200.equiv nb030nf_ntia, 208.equiv nb030nf_innerk, 212.equiv nb030nf_n, 216.equiv nb030nf_nn1, 220.equiv nb030nf_ntype, 224.equiv nb030nf_nouter, 228.equiv nb030nf_ninner, 232 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 248 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb030nf_nouter], eax mov [rsp + nb030nf_ninner], eax mov edi, [rdi] mov [rsp + nb030nf_nri], edi mov [rsp + nb030nf_iinr], rsi mov [rsp + nb030nf_jindex], rdx mov [rsp + nb030nf_jjnr], rcx mov [rsp + nb030nf_shift], r8 mov [rsp + nb030nf_shiftvec], r9 mov rdi, [rbp + nb030nf_p_ntype] mov edi, [rdi] mov [rsp + nb030nf_ntype], edi mov rax, [rbp + nb030nf_p_tabscale] movsd xmm3, [rax] shufpd xmm3, xmm3, 0 movapd [rsp + nb030nf_tsc], xmm3 ;# create constant floating-point factors on stack ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb030nf_half], eax mov [rsp + nb030nf_half+4], ebx movsd xmm1, [rsp + nb030nf_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 + nb030nf_half], xmm1 movapd [rsp + nb030nf_three], xmm3.nb030nf_threadloop: mov rsi, [rbp + nb030nf_count] ;# pointer to sync counter mov eax, [rsi].nb030nf_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 .nb030nf_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb030nf_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 + nb030nf_n], eax mov [rsp + nb030nf_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb030nf_outerstart jmp .nb030nf_end.nb030nf_outerstart:⌨️ 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?