nb_kernel100_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 970 行 · 第 1/2 页
S
970 行
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 + nb100_n] ;# get group index for i particle mov rdx, [rbp + nb100_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total coulomb energy and update it movhlps xmm6, xmm12 addsd xmm12, xmm6 ;# low xmm12 have the sum now ;# add earlier value from mem mov rax, [rbp + nb100_Vc] addsd xmm12, [rax + rdx*8] ;# move back to mem movsd [rax + rdx*8], xmm12 ;# finish if last mov ecx, [rsp + nb100_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb100_outerend ;# not last, iterate outer loop once more! mov [rsp + nb100_n], esi jmp .nb100_outer.nb100_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb100_nri] ;# esi already loaded with n above sub ecx, esi jz .nb100_end ;# non-zero, do one more workunit jmp .nb100_threadloop.nb100_end: mov eax, [rsp + nb100_nouter] mov ebx, [rsp + nb100_ninner] mov rcx, [rbp + nb100_outeriter] mov rdx, [rbp + nb100_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 312 emms pop r15 pop r14 pop r13 pop r12 pop rbx pop rbp ret.globl nb_kernel100nf_x86_64_sse2.globl _nb_kernel100nf_x86_64_sse2nb_kernel100nf_x86_64_sse2: _nb_kernel100nf_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb100nf_fshift, 16.equiv nb100nf_gid, 24.equiv nb100nf_pos, 32.equiv nb100nf_faction, 40.equiv nb100nf_charge, 48.equiv nb100nf_p_facel, 56.equiv nb100nf_argkrf, 64.equiv nb100nf_argcrf, 72.equiv nb100nf_Vc, 80.equiv nb100nf_type, 88.equiv nb100nf_p_ntype, 96.equiv nb100nf_vdwparam, 104.equiv nb100nf_Vvdw, 112.equiv nb100nf_p_tabscale, 120.equiv nb100nf_VFtab, 128.equiv nb100nf_invsqrta, 136.equiv nb100nf_dvda, 144.equiv nb100nf_p_gbtabscale, 152.equiv nb100nf_GBtab, 160.equiv nb100nf_p_nthreads, 168.equiv nb100nf_count, 176.equiv nb100nf_mtx, 184.equiv nb100nf_outeriter, 192.equiv nb100nf_inneriter, 200.equiv nb100nf_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb100nf_ix, 0.equiv nb100nf_iy, 16.equiv nb100nf_iz, 32.equiv nb100nf_iq, 48.equiv nb100nf_vctot, 64.equiv nb100nf_half, 80.equiv nb100nf_three, 96.equiv nb100nf_is3, 112.equiv nb100nf_ii3, 116.equiv nb100nf_nri, 120.equiv nb100nf_iinr, 128.equiv nb100nf_jindex, 136.equiv nb100nf_jjnr, 144.equiv nb100nf_shift, 156.equiv nb100nf_shiftvec, 164.equiv nb100nf_facel, 172.equiv nb100nf_innerjjnr, 180.equiv nb100nf_innerk, 188.equiv nb100nf_n, 192.equiv nb100nf_nn1, 196.equiv nb100nf_nouter, 200.equiv nb100nf_ninner, 204 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 216 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb100nf_nouter], eax mov [rsp + nb100nf_ninner], eax mov edi, [rdi] mov [rsp + nb100nf_nri], edi mov [rsp + nb100nf_iinr], rsi mov [rsp + nb100nf_jindex], rdx mov [rsp + nb100nf_jjnr], rcx mov [rsp + nb100nf_shift], r8 mov [rsp + nb100nf_shiftvec], r9 mov rsi, [rbp + nb100nf_p_facel] movsd xmm0, [rsi] movsd [rsp + nb100nf_facel], xmm0 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb100nf_half], eax mov [rsp + nb100nf_half+4], ebx movsd xmm1, [rsp + nb100nf_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 + nb100nf_half], xmm1 movapd [rsp + nb100nf_three], xmm3.nb100nf_threadloop: mov rsi, [rbp + nb100nf_count] ;# pointer to sync counter mov eax, [rsi].nb100nf_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 .nb100nf_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb100nf_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 + nb100nf_n], eax mov [rsp + nb100nf_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb100nf_outerstart jmp .nb100nf_end.nb100nf_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb100nf_nouter] mov [rsp + nb100nf_nouter], ebx.nb100nf_outer: mov rax, [rsp + nb100nf_shift] ;# rax = pointer into shift[] mov ebx, [rax + rsi*4] ;# rbx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov rax, [rsp + nb100nf_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 + nb100nf_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx+rsi*4] ;# ebx =ii mov rdx, [rbp + nb100nf_charge] movsd xmm3, [rdx + rbx*8] mulsd xmm3, [rsp + nb100nf_facel] shufpd xmm3, xmm3, 0 lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb100nf_pos] ;# rax = base of pos[] addsd xmm0, [rax + rbx*8] addsd xmm1, [rax + rbx*8 + 8] addsd xmm2, [rax + rbx*8 + 16] movapd [rsp + nb100nf_iq], xmm3 shufpd xmm0, xmm0, 0 shufpd xmm1, xmm1, 0 shufpd xmm2, xmm2, 0 movapd [rsp + nb100nf_ix], xmm0 movapd [rsp + nb100nf_iy], xmm1 movapd [rsp + nb100nf_iz], xmm2 mov [rsp + nb100nf_ii3], ebx ;# clear vctot xorpd xmm4, xmm4 movapd [rsp + nb100nf_vctot], xmm4 mov rax, [rsp + nb100nf_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 + nb100nf_pos] mov rax, [rsp + nb100nf_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb100nf_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 2 add ecx, [rsp + nb100nf_ninner] mov [rsp + nb100nf_ninner], ecx add edx, 0 mov [rsp + nb100nf_innerk], edx ;# number of innerloop atoms jge .nb100nf_unroll_loop jmp .nb100nf_checksingle.nb100nf_unroll_loop: ;# twice unrolled innerloop here mov rdx, [rsp + nb100nf_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] add qword ptr [rsp + nb100nf_innerjjnr], 8 ;# advance pointer (unrolled 2) mov rsi, [rbp + nb100nf_charge] ;# base of charge[] movlpd xmm3, [rsi + rax*8] ;# jq A movhpd xmm3, [rsi + rbx*8] ;# jq B movapd xmm5, [rsp + nb100nf_iq] mulpd xmm3, xmm5 ;# qq mov rsi, [rbp + nb100nf_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 lea rbx, [rbx + rbx*2] ;# move two coordinates to xmm0-xmm2 movlpd xmm0, [rsi + rax*8] movlpd xmm1, [rsi + rax*8 + 8] movlpd xmm2, [rsi + rax*8 + 16] movhpd xmm0, [rsi + rbx*8] movhpd xmm1, [rsi + rbx*8 + 8] movhpd xmm2, [rsi + rbx*8 + 16] ;# move nb100nf_ix-iz to xmm4-xmm6 movapd xmm4, [rsp + nb100nf_ix] movapd xmm5, [rsp + nb100nf_iy] movapd xmm6, [rsp + nb100nf_iz] ;# calc dr subpd xmm4, xmm0 subpd xmm5, xmm1 subpd xmm6, xmm2 ;# square it mulpd xmm4,xmm4 mulpd xmm5,xmm5 mulpd xmm6,xmm6 addpd xmm4, xmm5 addpd xmm4, xmm6 ;# rsq in xmm4 cvtpd2ps xmm5, xmm4 rsqrtps xmm5, xmm5 cvtps2pd xmm2, xmm5 ;# lu in low xmm2 ;# lookup seed in xmm2 movapd xmm5, xmm2 ;# copy of lu mulpd xmm2, xmm2 ;# lu*lu movapd xmm1, [rsp + nb100nf_three] mulpd xmm2, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb100nf_half] subpd xmm1, xmm2 ;# 30-rsq*lu*lu mulpd xmm1, xmm5 mulpd xmm1, xmm0 ;# xmm0=iter1 of rinv (new lu) movapd xmm5, xmm1 ;# copy of lu mulpd xmm1, xmm1 ;# lu*lu movapd xmm2, [rsp + nb100nf_three] mulpd xmm1, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb100nf_half] subpd xmm2, xmm1 ;# 30-rsq*lu*lu mulpd xmm2, xmm5 mulpd xmm0, xmm2 ;# xmm0=iter2 of rinv (new lu) movapd xmm5, [rsp + nb100nf_vctot] mulpd xmm3, xmm0 ;# xmm3=vcoul addpd xmm5, xmm3 movapd [rsp + nb100nf_vctot], xmm5 ;# should we do one more iteration? sub dword ptr [rsp + nb100nf_innerk], 2 jl .nb100nf_checksingle jmp .nb100nf_unroll_loop.nb100nf_checksingle: mov edx, [rsp + nb100nf_innerk] and edx, 1 jnz .nb100nf_dosingle jmp .nb100nf_updateouterdata.nb100nf_dosingle: mov rsi, [rbp + nb100nf_charge] mov rdi, [rbp + nb100nf_pos] mov rdx, [rsp + nb100nf_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] xorpd xmm3, xmm3 movsd xmm3, [rsi + rax*8] ;# jq A movapd xmm5, [rsp + nb100nf_iq] unpcklpd xmm3, xmm6 mulpd xmm3, xmm5 ;# qq mov rsi, [rbp + nb100nf_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 ;# move two coordinates to xmm0-xmm2 movlpd xmm0, [rsi + rax*8] movlpd xmm1, [rsi + rax*8 + 8] movlpd xmm2, [rsi + rax*8 + 16] ;# move nb100nf_ix-iz to xmm4-xmm6 movapd xmm4, [rsp + nb100nf_ix] movapd xmm5, [rsp + nb100nf_iy] movapd xmm6, [rsp + nb100nf_iz] ;# calc dr subsd xmm4, xmm0 subsd xmm5, xmm1 subsd xmm6, xmm2 ;# 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 + nb100nf_three] mulsd xmm2, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb100nf_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 + nb100nf_three] mulsd xmm1, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb100nf_half] subsd xmm2, xmm1 ;# 30-rsq*lu*lu mulsd xmm2, xmm5 mulsd xmm0, xmm2 ;# xmm0=iter2 of rinv (new lu) movlpd xmm5, [rsp + nb100nf_vctot] mulsd xmm3, xmm0 ;# xmm3=vcoul addsd xmm5, xmm3 movlpd [rsp + nb100nf_vctot], xmm5.nb100nf_updateouterdata: ;# get n from stack mov esi, [rsp + nb100nf_n] ;# get group index for i particle mov rdx, [rbp + nb100nf_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movapd xmm7, [rsp + nb100nf_vctot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# low xmm7 has the sum now ;# add earlier value from mem mov rax, [rbp + nb100nf_Vc] addsd xmm7, [rax + rdx*8] ;# move back to mem movsd [rax + rdx*8], xmm7 ;# finish if last mov ecx, [rsp + nb100nf_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb100nf_outerend ;# not last, iterate outer loop once more! mov [rsp + nb100nf_n], esi jmp .nb100nf_outer.nb100nf_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb100nf_nri] ;# esi already loaded with n above sub ecx, esi jz .nb100nf_end ;# non-zero, do one more workunit jmp .nb100nf_threadloop.nb100nf_end: mov eax, [rsp + nb100nf_nouter] mov ebx, [rsp + nb100nf_ninner] mov rcx, [rbp + nb100nf_outeriter] mov rdx, [rbp + nb100nf_inneriter] mov [rcx], eax mov [rdx], ebx add rsp, 216 emms pop r15 pop r14 pop r13 pop r12 pop rbx pop rbp ret⌨️ 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?