nb_kernel133_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,375 行 · 第 1/5 页
S
2,375 行
pop r14 pop r13 pop r12 pop rbx pop rbp ret.globl nb_kernel133nf_x86_64_sse2.globl _nb_kernel133nf_x86_64_sse2nb_kernel133nf_x86_64_sse2: _nb_kernel133nf_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb133nf_fshift, 16.equiv nb133nf_gid, 24.equiv nb133nf_pos, 32.equiv nb133nf_faction, 40.equiv nb133nf_charge, 48.equiv nb133nf_p_facel, 56.equiv nb133nf_argkrf, 64.equiv nb133nf_argcrf, 72.equiv nb133nf_Vc, 80.equiv nb133nf_type, 88.equiv nb133nf_p_ntype, 96.equiv nb133nf_vdwparam, 104.equiv nb133nf_Vvdw, 112.equiv nb133nf_p_tabscale, 120.equiv nb133nf_VFtab, 128.equiv nb133nf_invsqrta, 136.equiv nb133nf_dvda, 144.equiv nb133nf_p_gbtabscale, 152.equiv nb133nf_GBtab, 160.equiv nb133nf_p_nthreads, 168.equiv nb133nf_count, 176.equiv nb133nf_mtx, 184.equiv nb133nf_outeriter, 192.equiv nb133nf_inneriter, 200.equiv nb133nf_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse2 use .equiv nb133nf_ixO, 0.equiv nb133nf_iyO, 16.equiv nb133nf_izO, 32.equiv nb133nf_ixH1, 48.equiv nb133nf_iyH1, 64.equiv nb133nf_izH1, 80.equiv nb133nf_ixH2, 96.equiv nb133nf_iyH2, 112.equiv nb133nf_izH2, 128.equiv nb133nf_ixM, 144.equiv nb133nf_iyM, 160.equiv nb133nf_izM, 176.equiv nb133nf_iqH, 192.equiv nb133nf_iqM, 208.equiv nb133nf_qqH, 224.equiv nb133nf_qqM, 240.equiv nb133nf_c6, 256.equiv nb133nf_c12, 272.equiv nb133nf_vctot, 288.equiv nb133nf_Vvdwtot, 304.equiv nb133nf_half, 320.equiv nb133nf_three, 336.equiv nb133nf_tsc, 352.equiv nb133nf_rinvH1, 368.equiv nb133nf_rinvH2, 384.equiv nb133nf_rinvM, 400.equiv nb133nf_krsqH1, 416.equiv nb133nf_krsqH2, 432.equiv nb133nf_krsqM, 448.equiv nb133nf_rsqO, 464.equiv nb133nf_nri, 496.equiv nb133nf_iinr, 504.equiv nb133nf_jindex, 512.equiv nb133nf_jjnr, 520.equiv nb133nf_shift, 528.equiv nb133nf_shiftvec, 536.equiv nb133nf_facel, 544.equiv nb133nf_innerjjnr, 552.equiv nb133nf_is3, 560.equiv nb133nf_ii3, 564.equiv nb133nf_ntia, 568.equiv nb133nf_innerk, 572.equiv nb133nf_n, 576.equiv nb133nf_nn1, 580.equiv nb133nf_nouter, 584.equiv nb133nf_ninner, 588 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 600 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb133nf_nouter], eax mov [rsp + nb133nf_ninner], eax mov edi, [rdi] mov [rsp + nb133nf_nri], edi mov [rsp + nb133nf_iinr], rsi mov [rsp + nb133nf_jindex], rdx mov [rsp + nb133nf_jjnr], rcx mov [rsp + nb133nf_shift], r8 mov [rsp + nb133nf_shiftvec], r9 mov rsi, [rbp + nb133nf_p_facel] movsd xmm0, [rsi] movsd [rsp + nb133nf_facel], xmm0 mov rax, [rbp + nb133nf_p_tabscale] movsd xmm3, [rax] shufpd xmm3, xmm3, 0 movapd [rsp + nb133nf_tsc], xmm3 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb133nf_half], eax mov [rsp + nb133nf_half+4], ebx movsd xmm1, [rsp + nb133nf_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 + nb133nf_half], xmm1 movapd [rsp + nb133nf_three], xmm3 ;# assume we have at least one i particle - start directly mov rcx, [rsp + nb133nf_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx] ;# ebx =ii mov rdx, [rbp + nb133nf_charge] movsd xmm3, [rdx + rbx*8 + 8] movsd xmm4, [rdx + rbx*8 + 24] movsd xmm5, [rsp + nb133nf_facel] mulsd xmm3, xmm5 mulsd xmm4, xmm5 shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 movapd [rsp + nb133nf_iqH], xmm3 movapd [rsp + nb133nf_iqM], xmm4 mov rdx, [rbp + nb133nf_type] mov ecx, [rdx + rbx*4] shl ecx, 1 mov rdi, [rbp + nb133nf_p_ntype] imul ecx, [rdi] ;# rcx = ntia = 2*ntype*type[ii0] mov [rsp + nb133nf_ntia], ecx .nb133nf_threadloop: mov rsi, [rbp + nb133nf_count] ;# pointer to sync counter mov eax, [rsi].nb133nf_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 .nb133nf_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb133nf_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 + nb133nf_n], eax mov [rsp + nb133nf_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb133nf_outerstart jmp .nb133nf_end.nb133nf_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb133nf_nouter] mov [rsp + nb133nf_nouter], ebx.nb133nf_outer: mov rax, [rsp + nb133nf_shift] ;# eax = pointer into shift[] mov ebx, [rax+rsi*4] ;# ebx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb133nf_is3],ebx ;# store is3 mov rax, [rsp + nb133nf_shiftvec] ;# eax = base of shiftvec[] movsd xmm0, [rax + rbx*8] movsd xmm1, [rax + rbx*8 + 8] movsd xmm2, [rax + rbx*8 + 16] mov rcx, [rsp + nb133nf_iinr] ;# ecx = pointer into iinr[] mov ebx, [rcx+rsi*4] ;# ebx =ii movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 movapd xmm6, xmm0 movapd xmm7, xmm1 lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb133nf_pos] ;# eax = base of pos[] mov [rsp + nb133nf_ii3], ebx addsd xmm3, [rax + rbx*8] ;# ox addsd xmm4, [rax + rbx*8 + 8] ;# oy addsd xmm5, [rax + rbx*8 + 16] ;# oz addsd xmm6, [rax + rbx*8 + 24] ;# h1x addsd xmm7, [rax + rbx*8 + 32] ;# h1y shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 shufpd xmm6, xmm6, 0 shufpd xmm7, xmm7, 0 movapd [rsp + nb133nf_ixO], xmm3 movapd [rsp + nb133nf_iyO], xmm4 movapd [rsp + nb133nf_izO], xmm5 movapd [rsp + nb133nf_ixH1], xmm6 movapd [rsp + nb133nf_iyH1], xmm7 movsd xmm6, xmm2 movsd xmm3, xmm0 movsd xmm4, xmm1 movsd xmm5, xmm2 addsd xmm6, [rax + rbx*8 + 40] ;# h1z addsd xmm0, [rax + rbx*8 + 48] ;# h2x addsd xmm1, [rax + rbx*8 + 56] ;# h2y addsd xmm2, [rax + rbx*8 + 64] ;# h2z addsd xmm3, [rax + rbx*8 + 72] ;# mx addsd xmm4, [rax + rbx*8 + 80] ;# my addsd xmm5, [rax + rbx*8 + 88] ;# mz shufpd xmm6, xmm6, 0 shufpd xmm0, xmm0, 0 shufpd xmm1, xmm1, 0 shufpd xmm2, xmm2, 0 shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 movapd [rsp + nb133nf_izH1], xmm6 movapd [rsp + nb133nf_ixH2], xmm0 movapd [rsp + nb133nf_iyH2], xmm1 movapd [rsp + nb133nf_izH2], xmm2 movapd [rsp + nb133nf_ixM], xmm3 movapd [rsp + nb133nf_iyM], xmm4 movapd [rsp + nb133nf_izM], xmm5 ;# clear vctot xorpd xmm4, xmm4 movapd [rsp + nb133nf_vctot], xmm4 movapd [rsp + nb133nf_Vvdwtot], xmm4 mov rax, [rsp + nb133nf_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 + nb133nf_pos] mov rdi, [rbp + nb133nf_faction] mov rax, [rsp + nb133nf_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb133nf_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 2 add ecx, [rsp + nb133nf_ninner] mov [rsp + nb133nf_ninner], ecx add edx, 0 mov [rsp + nb133nf_innerk], edx ;# number of innerloop atoms jge .nb133nf_unroll_loop jmp .nb133nf_checksingle.nb133nf_unroll_loop: ;# twice unrolled innerloop here mov rdx, [rsp + nb133nf_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] add qword ptr [rsp + nb133nf_innerjjnr], 8 ;# advance pointer (unrolled 2) mov rsi, [rbp + nb133nf_charge] ;# base of charge[] movlpd xmm3, [rsi + rax*8] movhpd xmm3, [rsi + rbx*8] movapd xmm4, xmm3 mulpd xmm3, [rsp + nb133nf_iqM] mulpd xmm4, [rsp + nb133nf_iqH] movd mm0, eax ;# use mmx registers as temp storage movd mm1, ebx movapd [rsp + nb133nf_qqM], xmm3 movapd [rsp + nb133nf_qqH], xmm4 mov rsi, [rbp + nb133nf_type] mov eax, [rsi + rax*4] mov ebx, [rsi + rbx*4] mov rsi, [rbp + nb133nf_vdwparam] shl eax, 1 shl ebx, 1 mov edi, [rsp + nb133nf_ntia] add eax, edi add ebx, edi movlpd xmm6, [rsi + rax*8] ;# c6a movlpd xmm7, [rsi + rbx*8] ;# c6b movhpd xmm6, [rsi + rax*8 + 8] ;# c6a c12a movhpd xmm7, [rsi + rbx*8 + 8] ;# c6b c12b movapd xmm4, xmm6 unpcklpd xmm4, xmm7 unpckhpd xmm6, xmm7 movd eax, mm0 movd ebx, mm1 movapd [rsp + nb133nf_c6], xmm4 movapd [rsp + nb133nf_c12], xmm6 mov rsi, [rbp + nb133nf_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 ixO-izO to xmm4-xmm6 movapd xmm4, [rsp + nb133nf_ixO] movapd xmm5, [rsp + nb133nf_iyO] movapd xmm6, [rsp + nb133nf_izO] ;# 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 movapd xmm7, xmm4 ;# rsqO in xmm7 ;# move ixH1-izH1 to xmm4-xmm6 movapd xmm4, [rsp + nb133nf_ixH1] movapd xmm5, [rsp + nb133nf_iyH1] movapd xmm6, [rsp + nb133nf_izH1] ;# calc dr subpd xmm4, xmm0 subpd xmm5, xmm1 subpd xmm6, xmm2 ;# square it mulpd xmm4,xmm4 mulpd xmm5,xmm5 mulpd xmm6,xmm6 addpd xmm6, xmm5 addpd xmm6, xmm4 ;# rsqH1 in xmm6 ;# move ixH2-izH2 to xmm3-xmm5 movapd xmm3, [rsp + nb133nf_ixH2] movapd xmm4, [rsp + nb133nf_iyH2] movapd xmm5, [rsp + nb133nf_izH2] ;# calc dr subpd xmm3, xmm0 subpd xmm4, xmm1 subpd xmm5, xmm2 ;# square it mulpd xmm3,xmm3 mulpd xmm4,xmm4 mulpd xmm5,xmm5 addpd xmm5, xmm4 addpd xmm5, xmm3 ;# move ixM-izM to xmm2-xmm4 movapd xmm3, [rsp + nb133nf_iyM] movapd xmm4, [rsp + nb133nf_izM] subpd xmm3, xmm1 subpd xmm4, xmm2 movapd xmm2, [rsp + nb133nf_ixM] subpd xmm2, xmm0 ;# square it mulpd xmm2,xmm2 mulpd xmm3,xmm3 mulpd xmm4,xmm4 addpd xmm4, xmm3 addpd xmm4, xmm2 ;# rsqM in xmm4, rsqH2 in xmm5, rsqH1 in xmm6, rsqO in xmm7 movapd [rsp + nb133nf_rsqO], xmm7 ;# start with rsqH1 - put seed in xmm2 cvtpd2ps xmm2, xmm6 rsqrtps xmm2, xmm2 cvtps2pd xmm2, xmm2 movapd xmm3, xmm2 mulpd xmm2, xmm2 movapd xmm1, [rsp + nb133nf_three] mulpd xmm2, xmm6 ;# rsq*lu*lu subpd xmm1, xmm2 ;# 30-rsq*lu*lu mulpd xmm1, xmm3 ;# lu*(3-rsq*lu*lu) mulpd xmm1, [rsp + nb133nf_half] ;# iter1 ( new lu) movapd xmm3, xmm1 mulpd xmm1, xmm1 ;# lu*lu mulpd xmm6, xmm1 ;# rsq*lu*lu movapd xmm1, [rsp + nb133nf_three] subpd xmm1, xmm6 ;# 3-rsq*lu*lu mulpd xmm1, xmm3 ;# lu*( 3-rsq*lu*lu) mulpd xmm1, [rsp + nb133nf_half] ;# rinv movapd [rsp + nb133nf_rinvH1], xmm1 ;# rsqH2 - seed in xmm2 cvtpd2ps xmm2, xmm5 rsqrtps xmm2, xmm2 cvtps2pd xmm2, xmm2 movapd xmm3, xmm2 mulpd xmm2, xmm2 movapd xmm1, [rsp + nb133nf_three] mulpd xmm2, xmm5 ;# rsq*lu*lu subpd xmm1, xmm2 ;# 30-rsq*lu*lu mulpd xmm1, xmm3 ;# lu*(3-rsq*lu*lu) mulpd xmm1, [rsp + nb133nf_half] ;# iter1 ( new lu) movapd xmm3, xmm1 mulpd xmm1, xmm1 ;# lu*lu mulpd xmm5, xmm1 ;# rsq*lu*lu movapd xmm1, [rsp + nb133nf_three] subpd xmm1, xmm5 ;# 3-rsq*lu*lu mulpd xmm1, xmm3 ;# lu*( 3-rsq*lu*lu) mulpd xmm1, [rsp + nb133nf_half] ;# rinv movapd [rsp + nb133nf_rinvH2], xmm1 ;# rsqM - seed in xmm2 cvtpd2ps xmm2, xmm4 rsqrtps xmm2, xmm2 cvtps2pd xmm2, xmm2 movapd xmm3, xmm2 mulpd xmm2, xmm2 movapd xmm1, [rsp + nb133nf_three] mulpd xmm2, xmm4 ;# rsq*lu*lu subpd xmm1, xmm2 ;# 30-rsq*lu*lu mulpd xmm1, xmm3 ;# lu*(3-rsq*lu*lu) mulpd xmm1, [rsp + nb133nf_half] ;# iter1 ( new lu) movapd xmm3, xmm1 mulpd xmm1, xmm1 ;# lu*lu mulpd xmm4, xmm1 ;# rsq*lu*lu ⌨️ 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?