nb_kernel310_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,344 行 · 第 1/3 页
S
1,344 行
movapd [rsp + nb310_vctot], xmm5 mulpd xmm7, [rsp + nb310_tsc] subpd xmm0, xmm7 mulpd xmm0, xmm1 ;# fscal ;# calculate scalar force by multiplying dx/dy/dz with fscal mulpd xmm9, xmm0 mulpd xmm10, xmm0 mulpd xmm11, xmm0 addpd xmm3, xmm9 addpd xmm4, xmm10 addpd xmm6, xmm11 ;# now update f_i addpd xmm13, xmm9 addpd xmm14, xmm10 addpd xmm15, xmm11 movlpd [rdi + rax*8], xmm3 movlpd [rdi + rax*8 + 8], xmm4 movlpd [rdi + rax*8 + 16], xmm6 movhpd [rdi + rbx*8], xmm3 movhpd [rdi + rbx*8 + 8], xmm4 movhpd [rdi + rbx*8 + 16], xmm6 ;# should we do one more iteration? sub dword ptr [rsp + nb310_innerk], 2 jl .nb310_checksingle jmp .nb310_unroll_loop.nb310_checksingle: mov edx, [rsp + nb310_innerk] and edx, 1 jnz .nb310_dosingle jmp .nb310_updateouterdata.nb310_dosingle: mov rsi, [rbp + nb310_charge] mov rdi, [rbp + nb310_pos] mov rcx, [rsp + nb310_innerjjnr] mov eax, [rcx] mov rsi, [rbp + nb310_charge] ;# base of charge[] movsd xmm3, [rsi + rax*8] mulsd xmm3, [rsp + nb310_iq] movapd [rsp + nb310_qq], xmm3 mov rsi, [rbp + nb310_type] mov r8d, [rsi + rax*4] mov rsi, [rbp + nb310_vdwparam] shl r8d, 1 mov edi, [rsp + nb310_ntia] add r8d, edi movsd xmm4, [rsi + r8*8] movsd xmm6, [rsi + r8*8 + 8] movapd [rsp + nb310_c6], xmm4 movapd [rsp + nb310_c12], xmm6 mov rsi, [rbp + nb310_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 + nb310_ix] subsd xmm5, [rsp + nb310_iy] subsd xmm6, [rsp + nb310_iz] ;# store dr movapd xmm9, xmm4 movapd xmm10, xmm5 movapd xmm11, xmm6 ;# square it mulsd xmm4,xmm4 mulsd xmm5,xmm5 mulsd xmm6,xmm6 addsd xmm4, xmm5 addsd xmm4, xmm6 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 + nb310_three] mulsd xmm2, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb310_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 + nb310_three] mulsd xmm1, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb310_half] subsd xmm2, xmm1 ;# 30-rsq*lu*lu mulsd xmm2, xmm5 mulsd xmm0, xmm2 ;# xmm0=rinv movapd xmm1, xmm0 ;# xmm1=rinv mulsd xmm4, xmm0 ;# xmm4=r mulsd xmm4, [rsp + nb310_tsc] cvttsd2si r8d, xmm4 ;# mm6 = lu idx cvtsi2sd xmm5, r8d subsd xmm4, xmm5 movapd xmm3, xmm4 ;# xmm3=eps shl r8d, 2 ;# idx *= 4 mov rsi, [rbp + nb310_VFtab] movsd xmm4, [rsi + r8*8] movsd xmm5, [rsi + r8*8 + 8] movsd xmm6, [rsi + r8*8 + 16] movsd xmm7, [rsi + r8*8 + 24] movapd xmm0, xmm1 ;# rinv mulsd xmm0, xmm0 ;# rinvsq movapd xmm2, xmm0 ;# rinvsq mulsd xmm2, xmm2 ;# rinv4 mulsd xmm2, xmm0 ;# rinv6 movapd xmm12, xmm2 mulsd xmm12, xmm12 ;# rinv12 mulsd xmm2, [rsp + nb310_c6] ;# vvdw6=c6*rinv6 mulsd xmm12, [rsp + nb310_c12] ;# vvdw12=c12*rinv12 movapd xmm0, xmm12 subsd xmm12, xmm2 ;# Vvdw=Vvdw12-Vvdw6 ;# add potential to vvdwtot addsd xmm12, [rsp + nb310_Vvdwtot] movsd [rsp + nb310_Vvdwtot], xmm12 ;# coulomb table ready, in xmm4-xmm7 mulsd xmm7, xmm3 ;# Heps mulsd xmm6, xmm3 ;# Geps mulsd xmm7, xmm3 ;# Heps2 addsd xmm5, xmm6 ;# F+Geps addsd xmm5, xmm7 ;# F+Geps+Heps2 = Fp addsd xmm7, xmm7 ;# 2*Heps2 addsd xmm7, xmm6 ;# 2*Heps2+Geps addsd xmm7, xmm5 ;# FF = Fp + 2*Heps2 + Geps mulsd xmm5, xmm3 ;# eps*Fp addsd xmm5, xmm4 ;# VV mulsd xmm5, [rsp + nb310_qq] ;# VV*qq=vcoul mulsd xmm7, [rsp + nb310_qq] ;# FF*qq=fijC ;# LJ forces mulsd xmm2, [rsp + nb310_six] mulsd xmm0, [rsp + nb310_twelve] subsd xmm0, xmm2 mulsd xmm0, xmm1 ;# (12*vnb12-6*vnb6)*rinv ;# add potential to vctot addsd xmm5, [rsp + nb310_vctot] movsd [rsp + nb310_vctot], xmm5 mulsd xmm7, [rsp + nb310_tsc] subsd xmm0, xmm7 mulsd xmm0, xmm1 ;# fscal ;# calculate scalar force by multiplying dx/dy/dz with fscal mulsd xmm9, xmm0 mulsd xmm10, xmm0 mulsd xmm11, xmm0 ;# now update f_i addsd xmm13, xmm9 addsd xmm14, xmm10 addsd xmm15, xmm11 ;# the fj's - start by accumulating forces from memory mov rdi, [rbp + nb310_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 .nb310_updateouterdata: mov ecx, [rsp + nb310_ii3] mov rdi, [rbp + nb310_faction] mov rsi, [rbp + nb310_fshift] mov edx, [rsp + nb310_is3] ;# accumulate i forces in xmm13, xmm14, xmm15 movhlps xmm3, xmm13 movhlps xmm4, xmm14 movhlps xmm5, xmm15 addsd xmm13, xmm3 addsd xmm14, xmm4 addsd xmm15, xmm5 ;# sum is in low xmm13-xmm15 ;# increment i force movsd xmm3, [rdi + rcx*8] movsd xmm4, [rdi + rcx*8 + 8] movsd xmm5, [rdi + rcx*8 + 16] subsd xmm3, xmm13 subsd xmm4, xmm14 subsd xmm5, xmm15 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, xmm13 subsd xmm4, xmm14 subsd xmm5, xmm15 movsd [rsi + rdx*8], xmm3 movsd [rsi + rdx*8 + 8], xmm4 movsd [rsi + rdx*8 + 16], xmm5 ;# get n from stack mov esi, [rsp + nb310_n] ;# get group index for i particle mov rdx, [rbp + nb310_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movapd xmm7, [rsp + nb310_vctot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# low xmm7 has the sum now ;# add earlier value from mem mov rax, [rbp + nb310_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 + nb310_Vvdwtot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# low xmm7 has the sum now ;# add earlier value from mem mov rax, [rbp + nb310_Vvdw] addsd xmm7, [rax + rdx*8] ;# move back to mem movsd [rax + rdx*8], xmm7 ;# finish if last mov ecx, [rsp + nb310_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb310_outerend ;# not last, iterate outer loop once more! mov [rsp + nb310_n], esi jmp .nb310_outer.nb310_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb310_nri] ;# esi already loaded with n above sub ecx, esi jz .nb310_end ;# non-zero, do one more workunit jmp .nb310_threadloop.nb310_end: mov eax, [rsp + nb310_nouter] mov ebx, [rsp + nb310_ninner] mov rcx, [rbp + nb310_outeriter] mov rdx, [rbp + nb310_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 472 emms pop r15 pop r14 pop r13 pop r12 pop rbx pop rbp ret.globl nb_kernel310nf_x86_64_sse2.globl _nb_kernel310nf_x86_64_sse2nb_kernel310nf_x86_64_sse2: _nb_kernel310nf_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb310nf_fshift, 16.equiv nb310nf_gid, 24.equiv nb310nf_pos, 32.equiv nb310nf_faction, 40.equiv nb310nf_charge, 48.equiv nb310nf_p_facel, 56.equiv nb310nf_argkrf, 64.equiv nb310nf_argcrf, 72.equiv nb310nf_Vc, 80.equiv nb310nf_type, 88.equiv nb310nf_p_ntype, 96.equiv nb310nf_vdwparam, 104.equiv nb310nf_Vvdw, 112.equiv nb310nf_p_tabscale, 120.equiv nb310nf_VFtab, 128.equiv nb310nf_invsqrta, 136.equiv nb310nf_dvda, 144.equiv nb310nf_p_gbtabscale, 152.equiv nb310nf_GBtab, 160.equiv nb310nf_p_nthreads, 168.equiv nb310nf_count, 176.equiv nb310nf_mtx, 184.equiv nb310nf_outeriter, 192.equiv nb310nf_inneriter, 200.equiv nb310nf_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb310nf_ix, 0.equiv nb310nf_iy, 16.equiv nb310nf_iz, 32.equiv nb310nf_iq, 48.equiv nb310nf_tsc, 64.equiv nb310nf_qq, 80.equiv nb310nf_c6, 96.equiv nb310nf_c12, 112.equiv nb310nf_vctot, 128.equiv nb310nf_Vvdwtot, 144.equiv nb310nf_half, 160.equiv nb310nf_three, 176.equiv nb310nf_nri, 192.equiv nb310nf_iinr, 200.equiv nb310nf_jindex, 208.equiv nb310nf_jjnr, 216.equiv nb310nf_shift, 224.equiv nb310nf_shiftvec, 232.equiv nb310nf_facel, 240.equiv nb310nf_innerjjnr, 248.equiv nb310nf_is3, 256.equiv nb310nf_ii3, 260.equiv nb310nf_ntia, 264.equiv nb310nf_innerk, 268.equiv nb310nf_n, 272.equiv nb310nf_nn1, 276.equiv nb310nf_ntype, 280.equiv nb310nf_nouter, 284.equiv nb310nf_ninner, 288 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 312 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb310nf_nouter], eax mov [rsp + nb310nf_ninner], eax mov edi, [rdi] mov [rsp + nb310nf_nri], edi mov [rsp + nb310nf_iinr], rsi mov [rsp + nb310nf_jindex], rdx mov [rsp + nb310nf_jjnr], rcx mov [rsp + nb310nf_shift], r8 mov [rsp + nb310nf_shiftvec], r9 mov rdi, [rbp + nb310nf_p_ntype] mov edi, [rdi] mov [rsp + nb310nf_ntype], edi mov rsi, [rbp + nb310nf_p_facel] movsd xmm0, [rsi] movsd [rsp + nb310nf_facel], xmm0 mov rax, [rbp + nb310nf_p_tabscale] movsd xmm3, [rax] shufpd xmm3, xmm3, 0 movapd [rsp + nb310nf_tsc], xmm3 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb310nf_half], eax mov [rsp + nb310nf_half+4], ebx movsd xmm1, [rsp + nb310nf_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 + nb310nf_half], xmm1 movapd [rsp + nb310nf_three], xmm3.nb310nf_threadloop: mov rsi, [rbp + nb310nf_count] ;# pointer to sync counter mov eax, [rsi].nb310nf_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 .nb310nf_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb310nf_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 + nb310nf_n], eax mov [rsp + nb310nf_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb310nf_outerstart jmp .nb310nf_end.nb310nf_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb310nf_nouter] mov [rsp + nb310nf_nouter], ebx⌨️ 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?