nb_kernel202_ia32_sse2.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,125 行 · 第 1/5 页
S
2,125 行
;#;# $Id: nb_kernel202_ia32_sse2.intel_syntax.s,v 1.1.2.2 2007/11/28 18:35:34 spoel 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_kernel202_ia32_sse2.globl _nb_kernel202_ia32_sse2nb_kernel202_ia32_sse2: _nb_kernel202_ia32_sse2: .equiv nb202_p_nri, 8.equiv nb202_iinr, 12.equiv nb202_jindex, 16.equiv nb202_jjnr, 20.equiv nb202_shift, 24.equiv nb202_shiftvec, 28.equiv nb202_fshift, 32.equiv nb202_gid, 36.equiv nb202_pos, 40.equiv nb202_faction, 44.equiv nb202_charge, 48.equiv nb202_p_facel, 52.equiv nb202_argkrf, 56.equiv nb202_argcrf, 60.equiv nb202_Vc, 64.equiv nb202_type, 68.equiv nb202_p_ntype, 72.equiv nb202_vdwparam, 76.equiv nb202_Vvdw, 80.equiv nb202_p_tabscale, 84.equiv nb202_VFtab, 88.equiv nb202_invsqrta, 92.equiv nb202_dvda, 96.equiv nb202_p_gbtabscale, 100.equiv nb202_GBtab, 104.equiv nb202_p_nthreads, 108.equiv nb202_count, 112.equiv nb202_mtx, 116.equiv nb202_outeriter, 120.equiv nb202_inneriter, 124.equiv nb202_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse2 use .equiv nb202_ixO, 0.equiv nb202_iyO, 16.equiv nb202_izO, 32.equiv nb202_ixH1, 48.equiv nb202_iyH1, 64.equiv nb202_izH1, 80.equiv nb202_ixH2, 96.equiv nb202_iyH2, 112.equiv nb202_izH2, 128.equiv nb202_jxO, 144.equiv nb202_jyO, 160.equiv nb202_jzO, 176.equiv nb202_jxH1, 192.equiv nb202_jyH1, 208.equiv nb202_jzH1, 224.equiv nb202_jxH2, 240.equiv nb202_jyH2, 256.equiv nb202_jzH2, 272.equiv nb202_dxOO, 288.equiv nb202_dyOO, 304.equiv nb202_dzOO, 320.equiv nb202_dxOH1, 336.equiv nb202_dyOH1, 352.equiv nb202_dzOH1, 368.equiv nb202_dxOH2, 384.equiv nb202_dyOH2, 400.equiv nb202_dzOH2, 416.equiv nb202_dxH1O, 432.equiv nb202_dyH1O, 448.equiv nb202_dzH1O, 464.equiv nb202_dxH1H1, 480.equiv nb202_dyH1H1, 496.equiv nb202_dzH1H1, 512.equiv nb202_dxH1H2, 528.equiv nb202_dyH1H2, 544.equiv nb202_dzH1H2, 560.equiv nb202_dxH2O, 576.equiv nb202_dyH2O, 592.equiv nb202_dzH2O, 608.equiv nb202_dxH2H1, 624.equiv nb202_dyH2H1, 640.equiv nb202_dzH2H1, 656.equiv nb202_dxH2H2, 672.equiv nb202_dyH2H2, 688.equiv nb202_dzH2H2, 704.equiv nb202_qqOO, 720.equiv nb202_qqOH, 736.equiv nb202_qqHH, 752.equiv nb202_vctot, 768.equiv nb202_fixO, 784.equiv nb202_fiyO, 800.equiv nb202_fizO, 816.equiv nb202_fixH1, 832.equiv nb202_fiyH1, 848.equiv nb202_fizH1, 864.equiv nb202_fixH2, 880.equiv nb202_fiyH2, 896.equiv nb202_fizH2, 912.equiv nb202_fjxO, 928.equiv nb202_fjyO, 944.equiv nb202_fjzO, 960.equiv nb202_fjxH1, 976.equiv nb202_fjyH1, 992.equiv nb202_fjzH1, 1008.equiv nb202_fjxH2, 1024.equiv nb202_fjyH2, 1040.equiv nb202_fjzH2, 1056.equiv nb202_half, 1072.equiv nb202_three, 1088.equiv nb202_rsqOO, 1104.equiv nb202_rsqOH1, 1120.equiv nb202_rsqOH2, 1136.equiv nb202_rsqH1O, 1152.equiv nb202_rsqH1H1, 1168.equiv nb202_rsqH1H2, 1184.equiv nb202_rsqH2O, 1200.equiv nb202_rsqH2H1, 1216.equiv nb202_rsqH2H2, 1232.equiv nb202_rinvOO, 1248.equiv nb202_rinvOH1, 1264.equiv nb202_rinvOH2, 1280.equiv nb202_rinvH1O, 1296.equiv nb202_rinvH1H1, 1312.equiv nb202_rinvH1H2, 1328.equiv nb202_rinvH2O, 1344.equiv nb202_rinvH2H1, 1360.equiv nb202_rinvH2H2, 1376.equiv nb202_two, 1392.equiv nb202_krf, 1408.equiv nb202_crf, 1424.equiv nb202_is3, 1440.equiv nb202_ii3, 1444.equiv nb202_innerjjnr, 1448.equiv nb202_innerk, 1452.equiv nb202_n, 1456.equiv nb202_nn1, 1460.equiv nb202_nri, 1464.equiv nb202_nouter, 1468.equiv nb202_ninner, 1472.equiv nb202_salign, 1476 push ebp mov ebp,esp push eax push ebx push ecx push edx push esi push edi sub esp, 1480 ;# local stack space mov eax, esp and eax, 0xf sub esp, eax mov [esp + nb202_salign], eax emms ;# Move args passed by reference to stack mov ecx, [ebp + nb202_p_nri] mov ecx, [ecx] mov [esp + nb202_nri], ecx ;# zero iteration counters mov eax, 0 mov [esp + nb202_nouter], eax mov [esp + nb202_ninner], eax ;# create constant floating-point factors on stack mov eax, 0x00000000 ;# lower half of double 0.5 IEEE (hex) mov ebx, 0x3fe00000 mov [esp + nb202_half], eax mov [esp + nb202_half+4], ebx movsd xmm1, [esp + nb202_half] shufpd xmm1, xmm1, 0 ;# splat to all elements movapd xmm3, xmm1 addpd xmm3, xmm3 ;# 1.0 movapd xmm2, xmm3 addpd xmm2, xmm2 ;# 2.0 addpd xmm3, xmm2 ;# 3.0 movapd [esp + nb202_half], xmm1 movapd [esp + nb202_two], xmm2 movapd [esp + nb202_three], xmm3 mov esi, [ebp + nb202_argkrf] mov edi, [ebp + nb202_argcrf] movsd xmm5, [esi] movsd xmm6, [edi] shufpd xmm5, xmm5, 0 shufpd xmm6, xmm6, 0 movapd [esp + nb202_krf], xmm5 movapd [esp + nb202_crf], xmm6 ;# assume we have at least one i particle - start directly mov ecx, [ebp + nb202_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx] ;# ebx =ii mov edx, [ebp + nb202_charge] movsd xmm3, [edx + ebx*8] movsd xmm4, xmm3 movsd xmm5, [edx + ebx*8 + 8] mov esi, [ebp + nb202_p_facel] movsd xmm6, [esi] 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 [esp + nb202_qqOO], xmm3 movapd [esp + nb202_qqOH], xmm4 movapd [esp + nb202_qqHH], xmm5 .nb202_threadloop: mov esi, [ebp + nb202_count] ;# pointer to sync counter mov eax, [esi].nb202_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 .nb202_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [esp + nb202_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 [esp + nb202_n], eax mov [esp + nb202_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb202_outerstart jmp .nb202_end.nb202_outerstart: ;# ebx contains number of outer iterations add ebx, [esp + nb202_nouter] mov [esp + nb202_nouter], ebx.nb202_outer: mov eax, [ebp + nb202_shift] ;# eax = pointer into shift[] mov ebx, [eax + esi*4] ;# ebx=shift[n] lea ebx, [ebx + ebx*2] ;# ebx=3*is mov [esp + nb202_is3],ebx ;# store is3 mov eax, [ebp + nb202_shiftvec] ;# eax = base of shiftvec[] movsd xmm0, [eax + ebx*8] movsd xmm1, [eax + ebx*8 + 8] movsd xmm2, [eax + ebx*8 + 16] mov ecx, [ebp + nb202_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx+esi*4] ;# ebx =ii lea ebx, [ebx + ebx*2] ;# ebx = 3*ii=ii3 mov eax, [ebp + nb202_pos] ;# eax = base of pos[] mov [esp + nb202_ii3], ebx movapd xmm3, xmm0 movapd xmm4, xmm1 movapd xmm5, xmm2 addsd xmm3, [eax + ebx*8] addsd xmm4, [eax + ebx*8 + 8] addsd xmm5, [eax + ebx*8 + 16] shufpd xmm3, xmm3, 0 shufpd xmm4, xmm4, 0 shufpd xmm5, xmm5, 0 movapd [esp + nb202_ixO], xmm3 movapd [esp + nb202_iyO], xmm4 movapd [esp + nb202_izO], xmm5 movsd xmm3, xmm0 movsd xmm4, xmm1 movsd xmm5, xmm2 addsd xmm0, [eax + ebx*8 + 24] addsd xmm1, [eax + ebx*8 + 32] addsd xmm2, [eax + ebx*8 + 40] addsd xmm3, [eax + ebx*8 + 48] addsd xmm4, [eax + ebx*8 + 56] addsd xmm5, [eax + ebx*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 [esp + nb202_ixH1], xmm0 movapd [esp + nb202_iyH1], xmm1 movapd [esp + nb202_izH1], xmm2 movapd [esp + nb202_ixH2], xmm3 movapd [esp + nb202_iyH2], xmm4 movapd [esp + nb202_izH2], xmm5 ;# clear vctot and i forces xorpd xmm4, xmm4 movapd [esp + nb202_vctot], xmm4 movapd [esp + nb202_fixO], xmm4 movapd [esp + nb202_fiyO], xmm4 movapd [esp + nb202_fizO], xmm4 movapd [esp + nb202_fixH1], xmm4 movapd [esp + nb202_fiyH1], xmm4 movapd [esp + nb202_fizH1], xmm4 movapd [esp + nb202_fixH2], xmm4 movapd [esp + nb202_fiyH2], xmm4 movapd [esp + nb202_fizH2], xmm4 mov eax, [ebp + nb202_jindex] mov ecx, [eax + esi*4] ;# jindex[n] mov edx, [eax + esi*4 + 4] ;# jindex[n+1] sub edx, ecx ;# number of innerloop atoms mov esi, [ebp + nb202_pos] mov edi, [ebp + nb202_faction] mov eax, [ebp + nb202_jjnr] shl ecx, 2 add eax, ecx mov [esp + nb202_innerjjnr], eax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 2 add ecx, [esp + nb202_ninner] mov [esp + nb202_ninner], ecx add edx, 0 mov [esp + nb202_innerk], edx ;# number of innerloop atoms jge .nb202_unroll_loop jmp .nb202_checksingle.nb202_unroll_loop: ;# twice unrolled innerloop here mov edx, [esp + nb202_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] mov ebx, [edx + 4] add dword ptr [esp + nb202_innerjjnr], 8 ;# advance pointer (unrolled 2) mov esi, [ebp + nb202_pos] ;# base of pos[] lea eax, [eax + eax*2] ;# replace jnr with j3 lea ebx, [ebx + ebx*2] ;# move j coordinates to local temp variables movlpd xmm2, [esi + eax*8] movlpd xmm3, [esi + eax*8 + 8] movlpd xmm4, [esi + eax*8 + 16] movlpd xmm5, [esi + eax*8 + 24] movlpd xmm6, [esi + eax*8 + 32] movlpd xmm7, [esi + eax*8 + 40] movhpd xmm2, [esi + ebx*8] movhpd xmm3, [esi + ebx*8 + 8] movhpd xmm4, [esi + ebx*8 + 16] movhpd xmm5, [esi + ebx*8 + 24] movhpd xmm6, [esi + ebx*8 + 32] movhpd xmm7, [esi + ebx*8 + 40] movapd [esp + nb202_jxO], xmm2 movapd [esp + nb202_jyO], xmm3 movapd [esp + nb202_jzO], xmm4 movapd [esp + nb202_jxH1], xmm5 movapd [esp + nb202_jyH1], xmm6 movapd [esp + nb202_jzH1], xmm7 movlpd xmm2, [esi + eax*8 + 48] movlpd xmm3, [esi + eax*8 + 56] movlpd xmm4, [esi + eax*8 + 64] movhpd xmm2, [esi + ebx*8 + 48] movhpd xmm3, [esi + ebx*8 + 56] movhpd xmm4, [esi + ebx*8 + 64] movapd [esp + nb202_jxH2], xmm2 movapd [esp + nb202_jyH2], xmm3 movapd [esp + nb202_jzH2], xmm4 movapd xmm0, [esp + nb202_ixO] movapd xmm1, [esp + nb202_iyO] movapd xmm2, [esp + nb202_izO]⌨️ 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?