nb_kernel112_x86_64_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,338 行 · 第 1/5 页
S
2,338 行
;#;# $Id: nb_kernel112_x86_64_sse2.intel_syntax.s,v 1.1.2.2 2006/09/22 08:40:32 lindahl Exp $;#;# Gromacs 4.0 Copyright (c) 1991-2003 ;# David van der Spoel, Erik Lindahl;#;# This program is free software; you can redistribute it and/or;# modify it under the terms of the GNU General Public License;# as published by the Free Software Foundation; either version 2;# of the License, or (at your option) any later version.;#;# To help us fund GROMACS development, we humbly ask that you cite;# the research papers on the package. Check out http://www.gromacs.org;# ;# And Hey:;# Gnomes, ROck Monsters And Chili Sauce;#;# These files require GNU binutils 2.10 or later, since we;# use intel syntax for portability, or a recent version ;# of NASM that understands Extended 3DNow and SSE2 instructions.;# (NASM is normally only used with MS Visual C++).;# Since NASM and gnu as disagree on some definitions and use ;# completely different preprocessing options I have to introduce a;# trick: NASM uses ';' for comments, while gnu as uses '#' on x86.;# Gnu as treats ';' as a line break, i.e. ignores it. This is the;# reason why all comments need both symbols...;# The source is written for GNU as, with intel syntax. When you use;# NASM we redefine a couple of things. The false if-statement around ;# the following code is seen by GNU as, but NASM doesn't see it, so ;# the code inside is read by NASM but not gcc.; .if 0 # block below only read by NASM%define .section section%define .long dd%define .align align%define .globl global;# NASM only wants 'dword', not 'dword ptr'.%define ptr.equiv .equiv 2 %1 equ %2%endmacro; .endif # End of NASM-specific block; .intel_syntax noprefix # Line only read by gnu as .globl nb_kernel112_x86_64_sse2.globl _nb_kernel112_x86_64_sse2nb_kernel112_x86_64_sse2: _nb_kernel112_x86_64_sse2: ;# Room for return address and rbp (16 bytes).equiv nb112_fshift, 16.equiv nb112_gid, 24.equiv nb112_pos, 32.equiv nb112_faction, 40.equiv nb112_charge, 48.equiv nb112_p_facel, 56.equiv nb112_argkrf, 64.equiv nb112_argcrf, 72.equiv nb112_Vc, 80.equiv nb112_type, 88.equiv nb112_p_ntype, 96.equiv nb112_vdwparam, 104.equiv nb112_Vvdw, 112.equiv nb112_p_tabscale, 120.equiv nb112_VFtab, 128.equiv nb112_invsqrta, 136.equiv nb112_dvda, 144.equiv nb112_p_gbtabscale, 152.equiv nb112_GBtab, 160.equiv nb112_p_nthreads, 168.equiv nb112_count, 176.equiv nb112_mtx, 184.equiv nb112_outeriter, 192.equiv nb112_inneriter, 200.equiv nb112_work, 208 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse2 use .equiv nb112_ixO, 0.equiv nb112_iyO, 16.equiv nb112_izO, 32.equiv nb112_ixH1, 48.equiv nb112_iyH1, 64.equiv nb112_izH1, 80.equiv nb112_ixH2, 96.equiv nb112_iyH2, 112.equiv nb112_izH2, 128.equiv nb112_jxO, 144.equiv nb112_jyO, 160.equiv nb112_jzO, 176.equiv nb112_jxH1, 192.equiv nb112_jyH1, 208.equiv nb112_jzH1, 224.equiv nb112_jxH2, 240.equiv nb112_jyH2, 256.equiv nb112_jzH2, 272.equiv nb112_dxOO, 288.equiv nb112_dyOO, 304.equiv nb112_dzOO, 320.equiv nb112_dxOH1, 336.equiv nb112_dyOH1, 352.equiv nb112_dzOH1, 368.equiv nb112_dxOH2, 384.equiv nb112_dyOH2, 400.equiv nb112_dzOH2, 416.equiv nb112_dxH1O, 432.equiv nb112_dyH1O, 448.equiv nb112_dzH1O, 464.equiv nb112_dxH1H1, 480.equiv nb112_dyH1H1, 496.equiv nb112_dzH1H1, 512.equiv nb112_dxH1H2, 528.equiv nb112_dyH1H2, 544.equiv nb112_dzH1H2, 560.equiv nb112_dxH2O, 576.equiv nb112_dyH2O, 592.equiv nb112_dzH2O, 608.equiv nb112_dxH2H1, 624.equiv nb112_dyH2H1, 640.equiv nb112_dzH2H1, 656.equiv nb112_dxH2H2, 672.equiv nb112_dyH2H2, 688.equiv nb112_dzH2H2, 704.equiv nb112_qqOO, 720.equiv nb112_qqOH, 736.equiv nb112_qqHH, 752.equiv nb112_c6, 768.equiv nb112_c12, 784.equiv nb112_six, 800.equiv nb112_twelve, 816.equiv nb112_vctot, 832.equiv nb112_Vvdwtot, 848.equiv nb112_fixO, 864.equiv nb112_fiyO, 880.equiv nb112_fizO, 896.equiv nb112_fixH1, 912.equiv nb112_fiyH1, 928.equiv nb112_fizH1, 944.equiv nb112_fixH2, 960.equiv nb112_fiyH2, 976.equiv nb112_fizH2, 992.equiv nb112_fjxO, 1008.equiv nb112_fjyO, 1024.equiv nb112_fjzO, 1040.equiv nb112_fjxH1, 1056.equiv nb112_fjyH1, 1072.equiv nb112_fjzH1, 1088.equiv nb112_fjxH2, 1104.equiv nb112_fjyH2, 1120.equiv nb112_fjzH2, 1136.equiv nb112_half, 1152.equiv nb112_three, 1168.equiv nb112_rsqOO, 1184.equiv nb112_rsqOH1, 1200.equiv nb112_rsqOH2, 1216.equiv nb112_rsqH1O, 1232.equiv nb112_rsqH1H1, 1248.equiv nb112_rsqH1H2, 1264.equiv nb112_rsqH2O, 1280.equiv nb112_rsqH2H1, 1296.equiv nb112_rsqH2H2, 1312.equiv nb112_rinvOO, 1328.equiv nb112_rinvOH1, 1344.equiv nb112_rinvOH2, 1360.equiv nb112_rinvH1O, 1376.equiv nb112_rinvH1H1, 1392.equiv nb112_rinvH1H2, 1408.equiv nb112_rinvH2O, 1424.equiv nb112_rinvH2H1, 1440.equiv nb112_rinvH2H2, 1456.equiv nb112_is3, 1472.equiv nb112_ii3, 1476.equiv nb112_nri, 1492.equiv nb112_iinr, 1500.equiv nb112_jindex, 1508.equiv nb112_jjnr, 1516.equiv nb112_shift, 1524.equiv nb112_shiftvec, 1532.equiv nb112_facel, 1540.equiv nb112_innerjjnr, 1548.equiv nb112_innerk, 1556.equiv nb112_n, 1560.equiv nb112_nn1, 1564.equiv nb112_nouter, 1568.equiv nb112_ninner, 1572 push rbp mov rbp, rsp push rbx emms push r12 push r13 push r14 push r15 sub rsp, 1592 ;# local variable stack space (n*16+8) ;# zero 32-bit iteration counters mov eax, 0 mov [rsp + nb112_nouter], eax mov [rsp + nb112_ninner], eax mov edi, [rdi] mov [rsp + nb112_nri], edi mov [rsp + nb112_iinr], rsi mov [rsp + nb112_jindex], rdx mov [rsp + nb112_jjnr], rcx mov [rsp + nb112_shift], r8 mov [rsp + nb112_shiftvec], r9 mov rsi, [rbp + nb112_p_facel] movsd xmm0, [rsi] movsd [rsp + nb112_facel], xmm0 ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double half IEEE (hex) mov ebx, 0x3fe00000 mov [rsp + nb112_half], eax mov [rsp + nb112_half+4], ebx movsd xmm1, [rsp + nb112_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 xmm4, xmm3 addpd xmm4, xmm4 ;# six movapd xmm5, xmm4 addpd xmm5, xmm5 ;# twelve movapd [rsp + nb112_half], xmm1 movapd [rsp + nb112_three], xmm3 movapd [rsp + nb112_six], xmm4 movapd [rsp + nb112_twelve], xmm5 ;# assume we have at least one i particle - start directly mov rcx, [rsp + nb112_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx] ;# ebx =ii mov rdx, [rbp + nb112_charge] movsd xmm3, [rdx + rbx*8] movsd xmm4, xmm3 movsd xmm5, [rdx + rbx*8 + 8] mov rsi, [rbp + nb112_p_facel] movsd xmm0, [rsi] movsd xmm6, [rsp + nb112_facel] mulsd xmm3, xmm3 mulsd xmm4, xmm5 mulsd xmm5, xmm5 mulsd xmm3, xmm6 mulsd xmm4, xmm6 mulsd xmm5, xmm6 shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 movapd [rsp + nb112_qqOO], xmm3 movapd [rsp + nb112_qqOH], xmm4 movapd [rsp + nb112_qqHH], xmm5 xorpd xmm0, xmm0 mov rdx, [rbp + nb112_type] mov ecx, [rdx + rbx*4] shl ecx, 1 mov edx, ecx mov rdi, [rbp + nb112_p_ntype] imul ecx, [rdi] ;# rcx = ntia = 2*ntype*type[ii0] add edx, ecx mov rax, [rbp + nb112_vdwparam] movlpd xmm0, [rax + rdx*8] movlpd xmm1, [rax + rdx*8 + 8] shufpd xmm0, xmm0, 0 shufpd xmm1, xmm1, 0 movapd [rsp + nb112_c6], xmm0 movapd [rsp + nb112_c12], xmm1.nb112_threadloop: mov rsi, [rbp + nb112_count] ;# pointer to sync counter mov eax, [rsi].nb112_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 .nb112_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [rsp + nb112_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 + nb112_n], eax mov [rsp + nb112_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb112_outerstart jmp .nb112_end.nb112_outerstart: ;# ebx contains number of outer iterations add ebx, [rsp + nb112_nouter] mov [rsp + nb112_nouter], ebx.nb112_outer: mov rax, [rsp + nb112_shift] ;# rax = pointer into shift[] mov ebx, [rax+rsi*4] ;# rbx=shift[n] lea rbx, [rbx + rbx*2] ;# rbx=3*is mov [rsp + nb112_is3],ebx ;# store is3 mov rax, [rsp + nb112_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 + nb112_iinr] ;# rcx = pointer into iinr[] mov ebx, [rcx+rsi*4] ;# ebx =ii lea rbx, [rbx + rbx*2] ;# rbx = 3*ii=ii3 mov rax, [rbp + nb112_pos] ;# rax = base of pos[] mov [rsp + nb112_ii3], ebx movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 addsd xmm3, [rax + rbx*8] addsd xmm4, [rax + rbx*8 + 8] addsd xmm5, [rax + rbx*8 + 16] shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 movapd [rsp + nb112_ixO], xmm3 movapd [rsp + nb112_iyO], xmm4 movapd [rsp + nb112_izO], xmm5 movsd xmm3, xmm0 movsd xmm4, xmm1 movsd xmm5, xmm2 addsd xmm0, [rax + rbx*8 + 24] addsd xmm1, [rax + rbx*8 + 32] addsd xmm2, [rax + rbx*8 + 40] addsd xmm3, [rax + rbx*8 + 48] addsd xmm4, [rax + rbx*8 + 56] addsd xmm5, [rax + rbx*8 + 64] 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 + nb112_ixH1], xmm0 movapd [rsp + nb112_iyH1], xmm1 movapd [rsp + nb112_izH1], xmm2 movapd [rsp + nb112_ixH2], xmm3 movapd [rsp + nb112_iyH2], xmm4 movapd [rsp + nb112_izH2], xmm5 ;# clear vctot and i forces xorpd xmm4, xmm4 movapd [rsp + nb112_vctot], xmm4 movapd [rsp + nb112_Vvdwtot], xmm4 movapd [rsp + nb112_fixO], xmm4 movapd [rsp + nb112_fiyO], xmm4 movapd [rsp + nb112_fizO], xmm4 movapd [rsp + nb112_fixH1], xmm4 movapd [rsp + nb112_fiyH1], xmm4 movapd [rsp + nb112_fizH1], xmm4 movapd [rsp + nb112_fixH2], xmm4 movapd [rsp + nb112_fiyH2], xmm4 movapd [rsp + nb112_fizH2], xmm4 mov rax, [rsp + nb112_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 + nb112_pos] mov rdi, [rbp + nb112_faction] mov rax, [rsp + nb112_jjnr] shl ecx, 2 add rax, rcx mov [rsp + nb112_innerjjnr], rax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 2 add ecx, [rsp + nb112_ninner] mov [rsp + nb112_ninner], ecx add edx, 0 mov [rsp + nb112_innerk], edx ;# number of innerloop atoms jge .nb112_unroll_loop jmp .nb112_checksingle.nb112_unroll_loop: ;# twice unrolled innerloop here mov rdx, [rsp + nb112_innerjjnr] ;# pointer to jjnr[k] mov eax, [rdx] mov ebx, [rdx + 4] add qword ptr [rsp + nb112_innerjjnr], 8 ;# advance pointer (unrolled 2) mov rsi, [rbp + nb112_pos] ;# base of pos[] lea rax, [rax + rax*2] ;# replace jnr with j3 lea rbx, [rbx + rbx*2] ;# move j O coordinates to local temp variables 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] ;# xmm0 = Ox ;# xmm1 = Oy ;# xmm2 = Oz movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 movapd xmm6, xmm0 movapd xmm7, xmm1 movapd xmm8, xmm2 subpd xmm0, [rsp + nb112_ixO] subpd xmm1, [rsp + nb112_iyO] subpd xmm2, [rsp + nb112_izO] subpd xmm3, [rsp + nb112_ixH1] subpd xmm4, [rsp + nb112_iyH1] subpd xmm5, [rsp + nb112_izH1] subpd xmm6, [rsp + nb112_ixH2] subpd xmm7, [rsp + nb112_iyH2] subpd xmm8, [rsp + nb112_izH2] movapd [rsp + nb112_dxOO], xmm0 movapd [rsp + nb112_dyOO], xmm1 movapd [rsp + nb112_dzOO], xmm2 mulpd xmm0, xmm0 mulpd xmm1, xmm1 mulpd xmm2, xmm2 movapd [rsp + nb112_dxH1O], xmm3 movapd [rsp + nb112_dyH1O], xmm4 movapd [rsp + nb112_dzH1O], xmm5 mulpd xmm3, xmm3 mulpd xmm4, xmm4 mulpd xmm5, xmm5 movapd [rsp + nb112_dxH2O], xmm6 movapd [rsp + nb112_dyH2O], xmm7 movapd [rsp + nb112_dzH2O], xmm8 mulpd xmm6, xmm6 mulpd xmm7, xmm7 mulpd xmm8, xmm8 addpd xmm0, xmm1 addpd xmm0, xmm2 addpd xmm3, xmm4 addpd xmm3, xmm5 addpd xmm6, xmm7 addpd xmm6, xmm8 ;# start doing invsqrt for jO atoms⌨️ 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?