nb_kernel113_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,895 行 · 第 1/4 页
S
1,895 行
;#;# $Id: nb_kernel113_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_kernel113_ia32_sse.globl _nb_kernel113_ia32_ssenb_kernel113_ia32_sse: _nb_kernel113_ia32_sse: .equiv nb113_p_nri, 8.equiv nb113_iinr, 12.equiv nb113_jindex, 16.equiv nb113_jjnr, 20.equiv nb113_shift, 24.equiv nb113_shiftvec, 28.equiv nb113_fshift, 32.equiv nb113_gid, 36.equiv nb113_pos, 40.equiv nb113_faction, 44.equiv nb113_charge, 48.equiv nb113_p_facel, 52.equiv nb113_p_krf, 56.equiv nb113_p_crf, 60.equiv nb113_Vc, 64.equiv nb113_type, 68.equiv nb113_p_ntype, 72.equiv nb113_vdwparam, 76.equiv nb113_Vvdw, 80.equiv nb113_p_tabscale, 84.equiv nb113_VFtab, 88.equiv nb113_invsqrta, 92.equiv nb113_dvda, 96.equiv nb113_p_gbtabscale, 100.equiv nb113_GBtab, 104.equiv nb113_p_nthreads, 108.equiv nb113_count, 112.equiv nb113_mtx, 116.equiv nb113_outeriter, 120.equiv nb113_inneriter, 124.equiv nb113_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb113_ixO, 0.equiv nb113_iyO, 16.equiv nb113_izO, 32.equiv nb113_ixH1, 48.equiv nb113_iyH1, 64.equiv nb113_izH1, 80.equiv nb113_ixH2, 96.equiv nb113_iyH2, 112.equiv nb113_izH2, 128.equiv nb113_ixM, 144.equiv nb113_iyM, 160.equiv nb113_izM, 176.equiv nb113_iqM, 192.equiv nb113_iqH, 208.equiv nb113_dxO, 224.equiv nb113_dyO, 240.equiv nb113_dzO, 256.equiv nb113_dxH1, 272.equiv nb113_dyH1, 288.equiv nb113_dzH1, 304.equiv nb113_dxH2, 320.equiv nb113_dyH2, 336.equiv nb113_dzH2, 352.equiv nb113_dxM, 368.equiv nb113_dyM, 384.equiv nb113_dzM, 400.equiv nb113_qqM, 416.equiv nb113_qqH, 432.equiv nb113_rinvH1, 448.equiv nb113_rinvH2, 464.equiv nb113_rinvM, 480.equiv nb113_two, 496.equiv nb113_c6, 512.equiv nb113_c12, 528.equiv nb113_six, 544.equiv nb113_twelve, 560.equiv nb113_vctot, 576.equiv nb113_Vvdwtot, 592.equiv nb113_fixO, 608.equiv nb113_fiyO, 624.equiv nb113_fizO, 640.equiv nb113_fixH1, 656.equiv nb113_fiyH1, 672.equiv nb113_fizH1, 688.equiv nb113_fixH2, 704.equiv nb113_fiyH2, 720.equiv nb113_fizH2, 736.equiv nb113_fixM, 752.equiv nb113_fiyM, 768.equiv nb113_fizM, 784.equiv nb113_fjx, 800.equiv nb113_fjy, 816.equiv nb113_fjz, 832.equiv nb113_half, 848.equiv nb113_three, 864.equiv nb113_is3, 880.equiv nb113_ii3, 884.equiv nb113_ntia, 888.equiv nb113_innerjjnr, 892.equiv nb113_innerk, 896.equiv nb113_n, 900.equiv nb113_nn1, 904.equiv nb113_nri, 908.equiv nb113_nouter, 912.equiv nb113_ninner, 916.equiv nb113_salign, 920 push ebp mov ebp,esp push eax push ebx push ecx push edx push esi push edi sub esp, 924 ;# local stack space mov eax, esp and eax, 0xf sub esp, eax mov [esp + nb113_salign], eax emms ;# Move args passed by reference to stack mov ecx, [ebp + nb113_p_nri] mov ecx, [ecx] mov [esp + nb113_nri], ecx ;# zero iteration counters mov eax, 0 mov [esp + nb113_nouter], eax mov [esp + nb113_ninner], eax ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# constant 0.5 in IEEE (hex) mov [esp + nb113_half], eax movss xmm1, [esp + nb113_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 + nb113_half], xmm1 movaps [esp + nb113_two], xmm2 movaps [esp + nb113_three], xmm3 movaps [esp + nb113_six], xmm4 movaps [esp + nb113_twelve], xmm5 ;# assume we have at least one i particle - start directly mov ecx, [ebp + nb113_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx] ;# ebx =ii mov edx, [ebp + nb113_charge] movss xmm4, [edx + ebx*4 + 4] movss xmm3, [edx + ebx*4 + 12] mov esi, [ebp + nb113_p_facel] movss xmm5, [esi] mulss xmm3, xmm5 mulss xmm4, xmm5 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 movaps [esp + nb113_iqM], xmm3 movaps [esp + nb113_iqH], xmm4 mov edx, [ebp + nb113_type] mov ecx, [edx + ebx*4] shl ecx, 1 mov edi, [ebp + nb113_p_ntype] imul ecx, [edi] ;# ecx = ntia = 2*ntype*type[ii0] mov [esp + nb113_ntia], ecx .nb113_threadloop: mov esi, [ebp + nb113_count] ;# pointer to sync counter mov eax, [esi].nb113_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 .nb113_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [esp + nb113_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 + nb113_n], eax mov [esp + nb113_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb113_outerstart jmp .nb113_end .nb113_outerstart: ;# ebx contains number of outer iterations add ebx, [esp + nb113_nouter] mov [esp + nb113_nouter], ebx.nb113_outer: mov eax, [ebp + nb113_shift] ;# eax = pointer into shift[] mov ebx, [eax + esi*4] ;# ebx=shift[n] lea ebx, [ebx + ebx*2] ;# ebx=3*is mov [esp + nb113_is3],ebx ;# store is3 mov eax, [ebp + nb113_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 + nb113_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx + esi*4] ;# ebx =ii movaps xmm3, xmm0 movaps xmm4, xmm1 movaps xmm5, xmm2 movaps xmm6, xmm0 movaps xmm7, xmm1 lea ebx, [ebx + ebx*2] ;# ebx = 3*ii=ii3 mov eax, [ebp + nb113_pos] ;# eax = base of pos[] mov [esp + nb113_ii3], ebx addss xmm3, [eax + ebx*4] ;# ox addss xmm4, [eax + ebx*4 + 4] ;# oy addss xmm5, [eax + ebx*4 + 8] ;# oz addss xmm6, [eax + ebx*4 + 12] ;# h1x addss xmm7, [eax + ebx*4 + 16] ;# h1y shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 shufps xmm5, xmm5, 0 shufps xmm6, xmm6, 0 shufps xmm7, xmm7, 0 movaps [esp + nb113_ixO], xmm3 movaps [esp + nb113_iyO], xmm4 movaps [esp + nb113_izO], xmm5 movaps [esp + nb113_ixH1], xmm6 movaps [esp + nb113_iyH1], xmm7 movss xmm6, xmm2 movss xmm3, xmm0 movss xmm4, xmm1 movss xmm5, xmm2 addss xmm6, [eax + ebx*4 + 20] ;# h1z addss xmm0, [eax + ebx*4 + 24] ;# h2x addss xmm1, [eax + ebx*4 + 28] ;# h2y addss xmm2, [eax + ebx*4 + 32] ;# h2z addss xmm3, [eax + ebx*4 + 36] ;# mx addss xmm4, [eax + ebx*4 + 40] ;# my addss xmm5, [eax + ebx*4 + 44] ;# mz shufps xmm6, xmm6, 0 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 + nb113_izH1], xmm6 movaps [esp + nb113_ixH2], xmm0 movaps [esp + nb113_iyH2], xmm1 movaps [esp + nb113_izH2], xmm2 movaps [esp + nb113_ixM], xmm3 movaps [esp + nb113_iyM], xmm4 movaps [esp + nb113_izM], xmm5 ;# clear vctot and i forces xorps xmm4, xmm4 movaps [esp + nb113_vctot], xmm4 movaps [esp + nb113_Vvdwtot], xmm4 movaps [esp + nb113_fixO], xmm4 movaps [esp + nb113_fiyO], xmm4 movaps [esp + nb113_fizO], xmm4 movaps [esp + nb113_fixH1], xmm4 movaps [esp + nb113_fiyH1], xmm4 movaps [esp + nb113_fizH1], xmm4 movaps [esp + nb113_fixH2], xmm4 movaps [esp + nb113_fiyH2], xmm4 movaps [esp + nb113_fizH2], xmm4 movaps [esp + nb113_fixM], xmm4 movaps [esp + nb113_fiyM], xmm4 movaps [esp + nb113_fizM], xmm4 mov eax, [ebp + nb113_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 + nb113_pos] mov edi, [ebp + nb113_faction] mov eax, [ebp + nb113_jjnr] shl ecx, 2 add eax, ecx mov [esp + nb113_innerjjnr], eax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 4 add ecx, [esp + nb113_ninner] mov [esp + nb113_ninner], ecx add edx, 0 mov [esp + nb113_innerk], edx ;# number of innerloop atoms jge .nb113_unroll_loop jmp .nb113_odd_inner.nb113_unroll_loop: ;# quad-unroll innerloop here mov edx, [esp + nb113_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 + nb113_innerjjnr], 16 ;# advance pointer (unrolled 4) mov esi, [ebp + nb113_charge] ;# base of charge[] movss xmm3, [esi + eax*4] movss xmm4, [esi + ecx*4] movss xmm6, [esi + ebx*4] movss xmm7, [esi + edx*4] shufps xmm3, xmm6, 0 shufps xmm4, xmm7, 0 shufps xmm3, xmm4, 136 ;# constant 10001000 ;# all charges in xmm3 movaps xmm4, xmm3 ;# and in xmm4 mulps xmm3, [esp + nb113_iqM] mulps xmm4, [esp + nb113_iqH] movd mm0, eax ;# use mmx registers as temp storage movd mm1, ebx movd mm2, ecx movd mm3, edx movaps [esp + nb113_qqM], xmm3 movaps [esp + nb113_qqH], xmm4 mov esi, [ebp + nb113_type] mov eax, [esi + eax*4] mov ebx, [esi + ebx*4] mov ecx, [esi + ecx*4] mov edx, [esi + edx*4] mov esi, [ebp + nb113_vdwparam] shl eax, 1 shl ebx, 1 shl ecx, 1 shl edx, 1 mov edi, [esp + nb113_ntia] add eax, edi add ebx, edi add ecx, edi add edx, edi movlps xmm6, [esi + eax*4] movlps xmm7, [esi + ecx*4] movhps xmm6, [esi + ebx*4] movhps xmm7, [esi + edx*4] movaps xmm4, xmm6 shufps xmm4, xmm7, 136 ;# constant 10001000 shufps xmm6, xmm7, 221 ;# constant 11011101 movd eax, mm0 movd ebx, mm1 movd ecx, mm2 movd edx, mm3 movaps [esp + nb113_c6], xmm4 movaps [esp + nb113_c12], xmm6 mov esi, [ebp + nb113_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 four coordinates to xmm0-xmm2 movlps xmm4, [esi + eax*4] movlps xmm5, [esi + ecx*4] movss xmm2, [esi + eax*4 + 8] movss xmm6, [esi + ecx*4 + 8] movhps xmm4, [esi + ebx*4] movhps xmm5, [esi + edx*4] movss xmm0, [esi + ebx*4 + 8] movss xmm1, [esi + edx*4 + 8] shufps xmm2, xmm0, 0 shufps xmm6, xmm1, 0 movaps xmm0, xmm4 movaps xmm1, xmm4 shufps xmm2, xmm6, 136 ;# constant 10001000 shufps xmm0, xmm5, 136 ;# constant 10001000 shufps xmm1, xmm5, 221 ;# constant 11011101 ;# move ixO-izO to xmm4-xmm6 movaps xmm4, [esp + nb113_ixO] movaps xmm5, [esp + nb113_iyO] movaps xmm6, [esp + nb113_izO] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb113_dxO], xmm4 movaps [esp + nb113_dyO], xmm5 movaps [esp + nb113_dzO], xmm6 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm4, xmm5 addps xmm4, xmm6 movaps xmm7, xmm4⌨️ 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?