nb_kernel212_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,151 行 · 第 1/5 页
S
2,151 行
;#;# $Id: nb_kernel212_ia32_sse.intel_syntax.s,v 1.1.2.1 2006/03/01 15:18:29 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_kernel212_ia32_sse.globl _nb_kernel212_ia32_ssenb_kernel212_ia32_sse: _nb_kernel212_ia32_sse: .equiv nb212_p_nri, 8.equiv nb212_iinr, 12.equiv nb212_jindex, 16.equiv nb212_jjnr, 20.equiv nb212_shift, 24.equiv nb212_shiftvec, 28.equiv nb212_fshift, 32.equiv nb212_gid, 36.equiv nb212_pos, 40.equiv nb212_faction, 44.equiv nb212_charge, 48.equiv nb212_p_facel, 52.equiv nb212_argkrf, 56.equiv nb212_argcrf, 60.equiv nb212_Vc, 64.equiv nb212_type, 68.equiv nb212_p_ntype, 72.equiv nb212_vdwparam, 76.equiv nb212_Vvdw, 80.equiv nb212_p_tabscale, 84.equiv nb212_VFtab, 88.equiv nb212_invsqrta, 92.equiv nb212_dvda, 96.equiv nb212_p_gbtabscale, 100.equiv nb212_GBtab, 104.equiv nb212_p_nthreads, 108.equiv nb212_count, 112.equiv nb212_mtx, 116.equiv nb212_outeriter, 120.equiv nb212_inneriter, 124.equiv nb212_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb212_ixO, 0.equiv nb212_iyO, 16.equiv nb212_izO, 32.equiv nb212_ixH1, 48.equiv nb212_iyH1, 64.equiv nb212_izH1, 80.equiv nb212_ixH2, 96.equiv nb212_iyH2, 112.equiv nb212_izH2, 128.equiv nb212_jxO, 144.equiv nb212_jyO, 160.equiv nb212_jzO, 176.equiv nb212_jxH1, 192.equiv nb212_jyH1, 208.equiv nb212_jzH1, 224.equiv nb212_jxH2, 240.equiv nb212_jyH2, 256.equiv nb212_jzH2, 272.equiv nb212_dxOO, 288.equiv nb212_dyOO, 304.equiv nb212_dzOO, 320.equiv nb212_dxOH1, 336.equiv nb212_dyOH1, 352.equiv nb212_dzOH1, 368.equiv nb212_dxOH2, 384.equiv nb212_dyOH2, 400.equiv nb212_dzOH2, 416.equiv nb212_dxH1O, 432.equiv nb212_dyH1O, 448.equiv nb212_dzH1O, 464.equiv nb212_dxH1H1, 480.equiv nb212_dyH1H1, 496.equiv nb212_dzH1H1, 512.equiv nb212_dxH1H2, 528.equiv nb212_dyH1H2, 544.equiv nb212_dzH1H2, 560.equiv nb212_dxH2O, 576.equiv nb212_dyH2O, 592.equiv nb212_dzH2O, 608.equiv nb212_dxH2H1, 624.equiv nb212_dyH2H1, 640.equiv nb212_dzH2H1, 656.equiv nb212_dxH2H2, 672.equiv nb212_dyH2H2, 688.equiv nb212_dzH2H2, 704.equiv nb212_qqOO, 720.equiv nb212_qqOH, 736.equiv nb212_qqHH, 752.equiv nb212_c6, 768.equiv nb212_c12, 784.equiv nb212_six, 800.equiv nb212_twelve, 816.equiv nb212_vctot, 832.equiv nb212_Vvdwtot, 848.equiv nb212_fixO, 864.equiv nb212_fiyO, 880.equiv nb212_fizO, 896.equiv nb212_fixH1, 912.equiv nb212_fiyH1, 928.equiv nb212_fizH1, 944.equiv nb212_fixH2, 960.equiv nb212_fiyH2, 976.equiv nb212_fizH2, 992.equiv nb212_fjxO, 1008.equiv nb212_fjyO, 1024.equiv nb212_fjzO, 1040.equiv nb212_fjxH1, 1056.equiv nb212_fjyH1, 1072.equiv nb212_fjzH1, 1088.equiv nb212_fjxH2, 1104.equiv nb212_fjyH2, 1120.equiv nb212_fjzH2, 1136.equiv nb212_fjzH2b, 1140.equiv nb212_fjzH2c, 1144.equiv nb212_fjzH2d, 1148.equiv nb212_half, 1152.equiv nb212_three, 1168.equiv nb212_rsqOO, 1184.equiv nb212_rsqOH1, 1200.equiv nb212_rsqOH2, 1216.equiv nb212_rsqH1O, 1232.equiv nb212_rsqH1H1, 1248.equiv nb212_rsqH1H2, 1264.equiv nb212_rsqH2O, 1280.equiv nb212_rsqH2H1, 1296.equiv nb212_rsqH2H2, 1312.equiv nb212_rinvOO, 1328.equiv nb212_rinvOH1, 1344.equiv nb212_rinvOH2, 1360.equiv nb212_rinvH1O, 1376.equiv nb212_rinvH1H1, 1392.equiv nb212_rinvH1H2, 1408.equiv nb212_rinvH2O, 1424.equiv nb212_rinvH2H1, 1440.equiv nb212_rinvH2H2, 1456.equiv nb212_two, 1472.equiv nb212_krf, 1488.equiv nb212_crf, 1504.equiv nb212_is3, 1520.equiv nb212_ii3, 1524.equiv nb212_innerjjnr, 1528.equiv nb212_innerk, 1532.equiv nb212_n, 1536.equiv nb212_nn1, 1540.equiv nb212_nri, 1544.equiv nb212_nouter, 1548.equiv nb212_ninner, 1552.equiv nb212_salign, 1556 push ebp mov ebp,esp push eax push ebx push ecx push edx push esi push edi sub esp, 1560 ;# local stack space mov eax, esp and eax, 0xf sub esp, eax mov [esp + nb212_salign], eax emms ;# Move args passed by reference to stack mov ecx, [ebp + nb212_p_nri] mov ecx, [ecx] mov [esp + nb212_nri], ecx ;# zero iteration counters mov eax, 0 mov [esp + nb212_nouter], eax mov [esp + nb212_ninner], eax mov esi, [ebp + nb212_argkrf] mov edi, [ebp + nb212_argcrf] movss xmm5, [esi] movss xmm6, [edi] shufps xmm5, xmm5, 0 shufps xmm6, xmm6, 0 movaps [esp + nb212_krf], xmm5 movaps [esp + nb212_crf], xmm6 ;# assume we have at least one i particle - start directly mov ecx, [ebp + nb212_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx] ;# ebx =ii mov edx, [ebp + nb212_charge] movss xmm3, [edx + ebx*4] movss xmm4, xmm3 movss xmm5, [edx + ebx*4 + 4] mov esi, [ebp + nb212_p_facel] movss xmm6, [esi] mulss xmm3, xmm3 mulss xmm4, xmm5 mulss xmm5, xmm5 mulss xmm3, xmm6 mulss xmm4, xmm6 mulss xmm5, xmm6 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [esp + nb212_qqOO], xmm3 movaps [esp + nb212_qqOH], xmm4 movaps [esp + nb212_qqHH], xmm5 xorps xmm0, xmm0 mov edx, [ebp + nb212_type] mov ecx, [edx + ebx*4] shl ecx, 1 mov edx, ecx mov edi, [ebp + nb212_p_ntype] imul ecx, [edi] ;# ecx = ntia = 2*ntype*type[ii0] add edx, ecx mov eax, [ebp + nb212_vdwparam] movlps xmm0, [eax + edx*4] movaps xmm1, xmm0 shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 85 ;# constant 01010101 movaps [esp + nb212_c6], xmm0 movaps [esp + nb212_c12], xmm1 ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# constant 0.5 in IEEE (hex) mov [esp + nb212_half], eax movss xmm1, [esp + nb212_half] shufps xmm1, xmm1, 0 ;# splat to all elements movaps xmm2, xmm1 addps xmm2, xmm2 ;# constant 1.0 movaps xmm3, xmm2 addps xmm2, xmm2 ;# constant 2.0 addps xmm3, xmm2 ;# constant 3.0 movaps xmm4, xmm3 addps xmm4, xmm4 ;# 6.0 movaps xmm5, xmm4 addps xmm5, xmm5 ;# constant 12.0 movaps [esp + nb212_half], xmm1 movaps [esp + nb212_two], xmm2 movaps [esp + nb212_three], xmm3 movaps [esp + nb212_six], xmm4 movaps [esp + nb212_twelve], xmm5.nb212_threadloop: mov esi, [ebp + nb212_count] ;# pointer to sync counter mov eax, [esi].nb212_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 .nb212_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [esp + nb212_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 + nb212_n], eax mov [esp + nb212_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb212_outerstart jmp .nb212_end.nb212_outerstart: ;# ebx contains number of outer iterations add ebx, [esp + nb212_nouter] mov [esp + nb212_nouter], ebx.nb212_outer: mov eax, [ebp + nb212_shift] ;# eax = pointer into shift[] mov ebx, [eax + esi*4] ;# ebx=shift[n] lea ebx, [ebx + ebx*2] ;# ebx=3*is mov [esp + nb212_is3],ebx ;# store is3 mov eax, [ebp + nb212_shiftvec] ;# eax = base of shiftvec[] movss xmm0, [eax + ebx*4] movss xmm1, [eax + ebx*4 + 4] movss xmm2, [eax + ebx*4 + 8] mov ecx, [ebp + nb212_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx + esi*4] ;# ebx =ii lea ebx, [ebx + ebx*2] ;# ebx = 3*ii=ii3 mov eax, [ebp + nb212_pos] ;# eax = base of pos[] mov [esp + nb212_ii3], ebx movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 addss xmm3, [eax + ebx*4] addss xmm4, [eax + ebx*4 + 4] addss xmm5, [eax + ebx*4 + 8] shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [esp + nb212_ixO], xmm3 movaps [esp + nb212_iyO], xmm4 movaps [esp + nb212_izO], xmm5 movss xmm3, xmm0 movss xmm4, xmm1 movss xmm5, xmm2 addss xmm0, [eax + ebx*4 + 12] addss xmm1, [eax + ebx*4 + 16] addss xmm2, [eax + ebx*4 + 20] addss xmm3, [eax + ebx*4 + 24] addss xmm4, [eax + ebx*4 + 28] addss xmm5, [eax + ebx*4 + 32] shufps xmm0, xmm0, 0 shufps xmm1, xmm1, 0 shufps xmm2, xmm2, 0 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 movaps [esp + nb212_ixH1], xmm0 movaps [esp + nb212_iyH1], xmm1 movaps [esp + nb212_izH1], xmm2 movaps [esp + nb212_ixH2], xmm3 movaps [esp + nb212_iyH2], xmm4 movaps [esp + nb212_izH2], xmm5 ;# clear vctot and i forces xorps xmm4, xmm4 movaps [esp + nb212_vctot], xmm4 movaps [esp + nb212_Vvdwtot], xmm4 movaps [esp + nb212_fixO], xmm4 movaps [esp + nb212_fiyO], xmm4 movaps [esp + nb212_fizO], xmm4 movaps [esp + nb212_fixH1], xmm4 movaps [esp + nb212_fiyH1], xmm4 movaps [esp + nb212_fizH1], xmm4 movaps [esp + nb212_fixH2], xmm4 movaps [esp + nb212_fiyH2], xmm4 movaps [esp + nb212_fizH2], xmm4 mov eax, [ebp + nb212_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 + nb212_pos] mov edi, [ebp + nb212_faction] mov eax, [ebp + nb212_jjnr] shl ecx, 2 add eax, ecx mov [esp + nb212_innerjjnr], eax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 4 add ecx, [esp + nb212_ninner] mov [esp + nb212_ninner], ecx add edx, 0 mov [esp + nb212_innerk], edx ;# number of innerloop atoms jge .nb212_unroll_loop jmp .nb212_single_check.nb212_unroll_loop: ;# quad-unroll innerloop here mov edx, [esp + nb212_innerjjnr] ;# pointer to jjnr[k] mov eax, [edx] mov ebx, [edx + 4] mov ecx, [edx + 8] mov edx, [edx + 12] ;# eax-edx=jnr1-4 add dword ptr [esp + nb212_innerjjnr], 16 ;# advance pointer (unrolled 4) mov esi, [ebp + nb212_pos] ;# base of pos[] lea eax, [eax + eax*2] ;# replace jnr with j3 lea ebx, [ebx + ebx*2] lea ecx, [ecx + ecx*2] ;# replace jnr with j3 lea edx, [edx + edx*2] ;# move j coordinates to local temp variables movlps xmm2, [esi + eax*4] movlps xmm3, [esi + eax*4 + 12] movlps xmm4, [esi + eax*4 + 24]⌨️ 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?