nb_kernel102_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,198 行 · 第 1/5 页
S
2,198 行
;#;# $Id: nb_kernel102_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_kernel102_ia32_sse.globl _nb_kernel102_ia32_ssenb_kernel102_ia32_sse: _nb_kernel102_ia32_sse: .equiv nb102_p_nri, 8.equiv nb102_iinr, 12.equiv nb102_jindex, 16.equiv nb102_jjnr, 20.equiv nb102_shift, 24.equiv nb102_shiftvec, 28.equiv nb102_fshift, 32.equiv nb102_gid, 36.equiv nb102_pos, 40.equiv nb102_faction, 44.equiv nb102_charge, 48.equiv nb102_p_facel, 52.equiv nb102_p_krf, 56.equiv nb102_p_crf, 60.equiv nb102_Vc, 64.equiv nb102_type, 68.equiv nb102_p_ntype, 72.equiv nb102_vdwparam, 76.equiv nb102_Vvdw, 80.equiv nb102_p_tabscale, 84.equiv nb102_VFtab, 88.equiv nb102_invsqrta, 92.equiv nb102_dvda, 96.equiv nb102_p_gbtabscale, 100.equiv nb102_GBtab, 104.equiv nb102_p_nthreads, 108.equiv nb102_count, 112.equiv nb102_mtx, 116.equiv nb102_outeriter, 120.equiv nb102_inneriter, 124.equiv nb102_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb102_ixO, 0.equiv nb102_iyO, 16.equiv nb102_izO, 32.equiv nb102_ixH1, 48.equiv nb102_iyH1, 64.equiv nb102_izH1, 80.equiv nb102_ixH2, 96.equiv nb102_iyH2, 112.equiv nb102_izH2, 128.equiv nb102_jxO, 144.equiv nb102_jyO, 160.equiv nb102_jzO, 176.equiv nb102_jxH1, 192.equiv nb102_jyH1, 208.equiv nb102_jzH1, 224.equiv nb102_jxH2, 240.equiv nb102_jyH2, 256.equiv nb102_jzH2, 272.equiv nb102_dxOO, 288.equiv nb102_dyOO, 304.equiv nb102_dzOO, 320.equiv nb102_dxOH1, 336.equiv nb102_dyOH1, 352.equiv nb102_dzOH1, 368.equiv nb102_dxOH2, 384.equiv nb102_dyOH2, 400.equiv nb102_dzOH2, 416.equiv nb102_dxH1O, 432.equiv nb102_dyH1O, 448.equiv nb102_dzH1O, 464.equiv nb102_dxH1H1, 480.equiv nb102_dyH1H1, 496.equiv nb102_dzH1H1, 512.equiv nb102_dxH1H2, 528.equiv nb102_dyH1H2, 544.equiv nb102_dzH1H2, 560.equiv nb102_dxH2O, 576.equiv nb102_dyH2O, 592.equiv nb102_dzH2O, 608.equiv nb102_dxH2H1, 624.equiv nb102_dyH2H1, 640.equiv nb102_dzH2H1, 656.equiv nb102_dxH2H2, 672.equiv nb102_dyH2H2, 688.equiv nb102_dzH2H2, 704.equiv nb102_qqOO, 720.equiv nb102_qqOH, 736.equiv nb102_qqHH, 752.equiv nb102_vctot, 768.equiv nb102_fixO, 784.equiv nb102_fiyO, 800.equiv nb102_fizO, 816.equiv nb102_fixH1, 832.equiv nb102_fiyH1, 848.equiv nb102_fizH1, 864.equiv nb102_fixH2, 880.equiv nb102_fiyH2, 896.equiv nb102_fizH2, 912.equiv nb102_fjxO, 928.equiv nb102_fjyO, 944.equiv nb102_fjzO, 960.equiv nb102_fjxH1, 976.equiv nb102_fjyH1, 992.equiv nb102_fjzH1, 1008.equiv nb102_fjxH2, 1024.equiv nb102_fjyH2, 1040.equiv nb102_fjzH2, 1056.equiv nb102_fjzH2b, 1060.equiv nb102_fjzH2c, 1064.equiv nb102_fjzH2d, 1068.equiv nb102_half, 1072.equiv nb102_three, 1088.equiv nb102_rsqOO, 1104.equiv nb102_rsqOH1, 1120.equiv nb102_rsqOH2, 1136.equiv nb102_rsqH1O, 1152.equiv nb102_rsqH1H1, 1168.equiv nb102_rsqH1H2, 1184.equiv nb102_rsqH2O, 1200.equiv nb102_rsqH2H1, 1216.equiv nb102_rsqH2H2, 1232.equiv nb102_rinvOO, 1248.equiv nb102_rinvOH1, 1264.equiv nb102_rinvOH2, 1280.equiv nb102_rinvH1O, 1296.equiv nb102_rinvH1H1, 1312.equiv nb102_rinvH1H2, 1328.equiv nb102_rinvH2O, 1344.equiv nb102_rinvH2H1, 1360.equiv nb102_rinvH2H2, 1376.equiv nb102_is3, 1392.equiv nb102_ii3, 1396.equiv nb102_innerjjnr, 1400.equiv nb102_innerk, 1404.equiv nb102_n, 1408.equiv nb102_nn1, 1412.equiv nb102_nri, 1416.equiv nb102_nouter, 1420.equiv nb102_ninner, 1424.equiv nb102_salign, 1428 push ebp mov ebp,esp push eax push ebx push ecx push edx push esi push edi sub esp, 1432 ;# local stack space mov eax, esp and eax, 0xf sub esp, eax mov [esp + nb102_salign], eax emms ;# Move args passed by reference to stack mov ecx, [ebp + nb102_p_nri] mov ecx, [ecx] mov [esp + nb102_nri], ecx ;# zero iteration counters mov eax, 0 mov [esp + nb102_nouter], eax mov [esp + nb102_ninner], eax ;# assume we have at least one i particle - start directly mov ecx, [ebp + nb102_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx] ;# ebx =ii mov edx, [ebp + nb102_charge] movss xmm3, [edx + ebx*4] movss xmm4, xmm3 movss xmm5, [edx + ebx*4 + 4] mov esi, [ebp + nb102_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 + nb102_qqOO], xmm3 movaps [esp + nb102_qqOH], xmm4 movaps [esp + nb102_qqHH], xmm5 ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# constant 0.5 in IEEE (hex) mov [esp + nb102_half], eax movss xmm1, [esp + nb102_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 [esp + nb102_half], xmm1 movaps [esp + nb102_three], xmm3.nb102_threadloop: mov esi, [ebp + nb102_count] ;# pointer to sync counter mov eax, [esi].nb102_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 .nb102_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [esp + nb102_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 + nb102_n], eax mov [esp + nb102_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb102_outerstart jmp .nb102_end .nb102_outerstart: ;# ebx contains number of outer iterations add ebx, [esp + nb102_nouter] mov [esp + nb102_nouter], ebx.nb102_outer: mov eax, [ebp + nb102_shift] ;# eax = pointer into shift[] mov ebx, [eax+esi*4] ;# ebx=shift[n] lea ebx, [ebx + ebx*2] ;# ebx=3*is mov [esp + nb102_is3],ebx ;# store is3 mov eax, [ebp + nb102_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 + nb102_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx+esi*4] ;# ebx =ii lea ebx, [ebx + ebx*2] ;# ebx = 3*ii=ii3 mov eax, [ebp + nb102_pos] ;# eax = base of pos[] mov [esp + nb102_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 + nb102_ixO], xmm3 movaps [esp + nb102_iyO], xmm4 movaps [esp + nb102_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 + nb102_ixH1], xmm0 movaps [esp + nb102_iyH1], xmm1 movaps [esp + nb102_izH1], xmm2 movaps [esp + nb102_ixH2], xmm3 movaps [esp + nb102_iyH2], xmm4 movaps [esp + nb102_izH2], xmm5 ;# clear vctot and i forces xorps xmm4, xmm4 movaps [esp + nb102_vctot], xmm4 movaps [esp + nb102_fixO], xmm4 movaps [esp + nb102_fiyO], xmm4 movaps [esp + nb102_fizO], xmm4 movaps [esp + nb102_fixH1], xmm4 movaps [esp + nb102_fiyH1], xmm4 movaps [esp + nb102_fizH1], xmm4 movaps [esp + nb102_fixH2], xmm4 movaps [esp + nb102_fiyH2], xmm4 movaps [esp + nb102_fizH2], xmm4 mov eax, [ebp + nb102_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 + nb102_pos] mov edi, [ebp + nb102_faction] mov eax, [ebp + nb102_jjnr] shl ecx, 2 add eax, ecx mov [esp + nb102_innerjjnr], eax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 4 add ecx, [esp + nb102_ninner] mov [esp + nb102_ninner], ecx add edx, 0 mov [esp + nb102_innerk], edx ;# number of innerloop atoms jge .nb102_unroll_loop jmp .nb102_single_check.nb102_unroll_loop: ;# quad-unroll innerloop here mov edx, [esp + nb102_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 + nb102_innerjjnr], 16 ;# advance pointer (unrolled 4) mov esi, [ebp + nb102_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] movlps xmm5, [esi + ebx*4] movlps xmm6, [esi + ebx*4 + 12] movlps xmm7, [esi + ebx*4 + 24] movhps xmm2, [esi + ecx*4] movhps xmm3, [esi + ecx*4 + 12] movhps xmm4, [esi + ecx*4 + 24] movhps xmm5, [esi + edx*4] movhps xmm6, [esi + edx*4 + 12] movhps xmm7, [esi + edx*4 + 24] ;# current state: ;# xmm2= jxOa jyOa jxOc jyOc ;# xmm3= jxH1a jyH1a jxH1c jyH1c ;# xmm4= jxH2a jyH2a jxH2c jyH2c ;# xmm5= jxOb jyOb jxOd jyOd ;# xmm6= jxH1b jyH1b jxH1d jyH1d ;# xmm7= jxH2b jyH2b jxH2d jyH2d movaps xmm0, xmm2 movaps xmm1, xmm3 unpcklps xmm0, xmm5 ;# xmm0= jxOa jxOb jyOa jyOb unpcklps xmm1, xmm6 ;# xmm1= jxH1a jxH1b jyH1a jyH1b unpckhps xmm2, xmm5 ;# xmm2= jxOc jxOd jyOc jyOd unpckhps xmm3, xmm6 ;# xmm3= jxH1c jxH1d jyH1c jyH1d movaps xmm5, xmm4 movaps xmm6, xmm0 unpcklps xmm4, xmm7 ;# xmm4= jxH2a jxH2b jyH2a jyH2b unpckhps xmm5, xmm7 ;# xmm5= jxH2c jxH2d jyH2c jyH2d movaps xmm7, xmm1 movlhps xmm0, xmm2 ;# xmm0= jxOa jxOb jxOc jxOd movaps [esp + nb102_jxO], xmm0 movhlps xmm2, xmm6 ;# xmm2= jyOa jyOb jyOc jyOd movaps [esp + nb102_jyO], xmm2 movlhps xmm1, xmm3 movaps [esp + nb102_jxH1], xmm1 movhlps xmm3, xmm7 movaps xmm6, xmm4 movaps [esp + nb102_jyH1], xmm3 movlhps xmm4, xmm5 movaps [esp + nb102_jxH2], xmm4 movhlps xmm5, xmm6 movaps [esp + nb102_jyH2], xmm5 movss xmm0, [esi + eax*4 + 8] movss xmm1, [esi + eax*4 + 20] movss xmm2, [esi + eax*4 + 32]⌨️ 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?