nb_kernel210_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,183 行 · 第 1/3 页
S
1,183 行
mulpd xmm4, xmm2 ;# xmm4=total fscal addpd xmm6, [rsp + nb210_vctot] ;# the fj's - start by accumulating forces from memory mov rdi, [rbp + nb210_faction] movlpd xmm0, [rdi + rax*8] movlpd xmm1, [rdi + rax*8 + 8] movlpd xmm2, [rdi + rax*8 + 16] movhpd xmm0, [rdi + rbx*8] movhpd xmm1, [rdi + rbx*8 + 8] movhpd xmm2, [rdi + rbx*8 + 16] movapd [rsp + nb210_vctot], xmm6 movapd [rsp + nb210_Vvdwtot], xmm5 mulpd xmm9, xmm4 mulpd xmm10, xmm4 mulpd xmm11, xmm4 addpd xmm0, xmm9 addpd xmm1, xmm10 addpd xmm2, xmm11 movlpd [rdi + rax*8], xmm0 movlpd [rdi + rax*8 + 8], xmm1 movlpd [rdi + rax*8 + 16], xmm2 ;# now update f_i addpd xmm13, xmm9 addpd xmm14, xmm10 addpd xmm15, xmm11 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 + nb210_innerk], 2 jl .nb210_checksingle jmp .nb210_unroll_loop.nb210_checksingle: mov edx, [rsp + nb210_innerk] and edx, 1 jnz .nb210_dosingle jmp .nb210_updateouterdata.nb210_dosingle: mov rsi, [rbp + nb210_charge] mov rdi, [rbp + nb210_pos] mov rcx, [rsp + nb210_innerjjnr] mov eax, [rcx] mov rsi, [rbp + nb210_charge] ;# base of charge[] movsd xmm3, [rsi + rax*8] mulsd xmm3, [rsp + nb210_iq] ;# qq mov rsi, [rbp + nb210_type] mov r8d, [rsi + rax*4] mov rsi, [rbp + nb210_vdwparam] shl r8d, 1 mov edi, [rsp + nb210_ntia] add r8d, edi movsd xmm4, [rsi + r8*8] movsd xmm6, [rsi + r8*8 + 8] movapd [rsp + nb210_c6], xmm4 movapd [rsp + nb210_c12], xmm6 mov rsi, [rbp + nb210_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 + nb210_ix] subsd xmm5, [rsp + nb210_iy] subsd xmm6, [rsp + nb210_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 ;# rsq in xmm4 cvtsd2ss xmm5, xmm4 rsqrtss xmm5, xmm5 cvtss2sd xmm2, xmm5 ;# lu in low xmm2 movapd xmm7, [rsp + nb210_krf] ;# lookup seed in xmm2 movapd xmm5, xmm2 ;# copy of lu mulsd xmm2, xmm2 ;# lu*lu movapd xmm1, [rsp + nb210_three] mulsd xmm7, xmm4 ;# krsq mulsd xmm2, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb210_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 + nb210_three] mulsd xmm1, xmm4 ;# rsq*lu*lu movapd xmm0, [rsp + nb210_half] subsd xmm2, xmm1 ;# 30-rsq*lu*lu mulsd xmm2, xmm5 mulsd xmm0, xmm2 ;# xmm0=rinv movapd xmm4, xmm0 mulsd xmm4, xmm4 ;# xmm4=rinvsq movapd xmm6, xmm0 addsd xmm6, xmm7 ;# xmm6=rinv+ krsq movapd xmm1, xmm4 subsd xmm6, [rsp + nb210_crf] mulsd xmm1, xmm4 mulsd xmm1, xmm4 ;# xmm1=rinvsix movapd xmm2, xmm1 mulsd xmm2, xmm2 ;# xmm2=rinvtwelve mulsd xmm6, xmm3 ;# xmm6=vcoul=qq*(rinv+ krsq) addsd xmm7, xmm7 mulsd xmm1, [rsp + nb210_c6] mulsd xmm2, [rsp + nb210_c12] movapd xmm5, xmm2 subsd xmm5, xmm1 ;# Vvdw=Vvdw12-Vvdw6 addsd xmm5, [rsp + nb210_Vvdwtot] mulsd xmm1, [rsp + nb210_six] mulsd xmm2, [rsp + nb210_twelve] subsd xmm2, xmm1 subsd xmm0, xmm7 mulsd xmm3, xmm0 addsd xmm2, xmm3 mulsd xmm4, xmm2 ;# xmm4=total fscal addsd xmm6, [rsp + nb210_vctot] movsd [rsp + nb210_vctot], xmm6 movsd [rsp + nb210_Vvdwtot], xmm5 mov rdi, [rbp + nb210_faction] mulsd xmm9, xmm4 mulsd xmm10, xmm4 mulsd xmm11, xmm4 ;# now update f_i addsd xmm13, xmm9 addsd xmm14, xmm10 addsd xmm15, xmm11 ;# the fj's - start by accumulating forces from memory 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 .nb210_updateouterdata: mov ecx, [rsp + nb210_ii3] mov rdi, [rbp + nb210_faction] mov rsi, [rbp + nb210_fshift] mov edx, [rsp + nb210_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 + nb210_n] ;# get group index for i particle mov rdx, [rbp + nb210_gid] ;# base of gid[] mov edx, [rdx + rsi*4] ;# ggid=gid[n] ;# accumulate total potential energy and update it movapd xmm7, [rsp + nb210_vctot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# low xmm7 has the sum now ;# add earlier value from mem mov rax, [rbp + nb210_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 + nb210_Vvdwtot] ;# accumulate movhlps xmm6, xmm7 addsd xmm7, xmm6 ;# low xmm7 has the sum now ;# add earlier value from mem mov rax, [rbp + nb210_Vvdw] addsd xmm7, [rax + rdx*8] ;# move back to mem movsd [rax + rdx*8], xmm7 ;# finish if last mov ecx, [rsp + nb210_nn1] ;# esi already loaded with n inc esi sub ecx, esi jz .nb210_outerend ;# not last, iterate outer loop once more! mov [rsp + nb210_n], esi jmp .nb210_outer.nb210_outerend: ;# check if more outer neighborlists remain mov ecx, [rsp + nb210_nri] ;# esi already loaded with n above sub ecx, esi jz .nb210_end ;# non-zero, do one more workunit jmp .nb210_threadloop.nb210_end: mov eax, [rsp + nb210_nouter] mov ebx, [rsp + nb210_ninner] mov rcx, [rbp + nb210_outeriter] mov rdx, [rbp + nb210_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_kernel210nf_x86_64_sse2.globl _nb_kernel210nf_x86_64_sse2nb_kernel210nf_x86_64_sse2: _nb_kernel210nf_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb210nf_fshift, 16.equiv nb210nf_gid, 24.equiv nb210nf_pos, 32.equiv nb210nf_faction, 40.equiv nb210nf_charge, 48.equiv nb210nf_p_facel, 56.equiv nb210nf_argkrf, 64.equiv nb210nf_argcrf, 72.equiv nb210nf_Vc, 80.equiv nb210nf_type, 88.equiv nb210nf_p_ntype, 96.equiv nb210nf_vdwparam, 104.equiv nb210nf_Vvdw, 112.equiv nb210nf_p_tabscale, 120.equiv nb210nf_VFtab, 128.equiv nb210nf_invsqrta, 136.equiv nb210nf_dvda, 144.equiv nb210nf_p_gbtabscale, 152.equiv nb210nf_GBtab, 160.equiv nb210nf_p_nthreads, 168.equiv nb210nf_count, 176.equiv nb210nf_mtx, 184.equiv nb210nf_outeriter, 192.equiv nb210nf_inneriter, 200.equiv nb210nf_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb210nf_ix, 0.equiv nb210nf_iy, 16.equiv nb210nf_iz, 32.equiv nb210nf_iq, 48.equiv nb210nf_c6, 64.equiv nb210nf_c12, 80.equiv nb210nf_vctot, 96.equiv nb210nf_Vvdwtot, 112.equiv nb210nf_half, 128.equiv nb210nf_three, 144.equiv nb210nf_krf, 160.equiv nb210nf_crf, 176.equiv nb210nf_nri, 192.equiv nb210nf_iinr, 200.equiv nb210nf_jindex, 208.equiv nb210nf_jjnr, 216.equiv nb210nf_shift, 224.equiv nb210nf_shiftvec, 232.equiv nb210nf_facel, 240.equiv nb210nf_innerjjnr, 248.equiv nb210nf_is3, 256.equiv nb210nf_ii3, 260.equiv nb210nf_ntia, 264.equiv nb210nf_innerk, 268.equiv nb210nf_n, 272.equiv nb210nf_nn1, 276.equiv nb210nf_ntype, 280.equiv nb210nf_nouter, 284.equiv nb210nf_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 + nb210nf_nouter], eax mov [rsp + nb210nf_ninner], eax mov edi, [rdi] mov [rsp + nb210nf_nri], edi mov [rsp + nb210nf_iinr], rsi mov [rsp + nb210nf_jindex], rdx mov [rsp + nb210nf_jjnr], rcx mov [rsp + nb210nf_shift], r8 mov [rsp + nb210nf_shiftvec], r9 mov rdi, [rbp + nb210nf_p_ntype] mov edi, [rdi] mov [rsp + nb210nf_ntype], edi mov rsi, [rbp + nb210nf_p_facel] movsd xmm0, [rsi] movsd [rsp + nb210nf_facel], xmm0 mov rsi, [rbp + nb210nf_argkrf] mov rdi, [rbp + nb210nf_argcrf] movsd xmm1, [rsi] movsd xmm2, [rdi] shufpd xmm1, xmm1, 0 shufpd xmm2, xmm2, 0 movapd [rsp + nb210nf_krf], xmm1 movapd [rsp + nb210nf_crf], xmm2 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb210nf_half], eax mov [rsp + nb210nf_half+4], ebx movsd xmm1, [rsp + nb210nf_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 + nb210nf_half], xmm1 movapd [rsp + nb210nf_three], xmm3.nb210nf_threadloop: mov rsi, [rbp + nb210nf_count] ;# pointer to sync counter mov eax, [rsi].nb210nf_spinlock: mov ebx, eax ;# ebx=*count=nn0 add ebx, 1 ;# ebx=nn1=nn0+10⌨️ 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?