nb_kernel333_ia32_sse.intel_syntax.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 2,386 行 · 第 1/5 页
S
2,386 行
;#;# $Id: nb_kernel333_ia32_sse.intel_syntax.s,v 1.1.2.1 2006/03/01 15:18:31 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_kernel333_ia32_sse.globl _nb_kernel333_ia32_ssenb_kernel333_ia32_sse: _nb_kernel333_ia32_sse: .equiv nb333_p_nri, 8.equiv nb333_iinr, 12.equiv nb333_jindex, 16.equiv nb333_jjnr, 20.equiv nb333_shift, 24.equiv nb333_shiftvec, 28.equiv nb333_fshift, 32.equiv nb333_gid, 36.equiv nb333_pos, 40.equiv nb333_faction, 44.equiv nb333_charge, 48.equiv nb333_p_facel, 52.equiv nb333_argkrf, 56.equiv nb333_argcrf, 60.equiv nb333_Vc, 64.equiv nb333_type, 68.equiv nb333_p_ntype, 72.equiv nb333_vdwparam, 76.equiv nb333_Vvdw, 80.equiv nb333_p_tabscale, 84.equiv nb333_VFtab, 88.equiv nb333_invsqrta, 92.equiv nb333_dvda, 96.equiv nb333_p_gbtabscale, 100.equiv nb333_GBtab, 104.equiv nb333_p_nthreads, 108.equiv nb333_count, 112.equiv nb333_mtx, 116.equiv nb333_outeriter, 120.equiv nb333_inneriter, 124.equiv nb333_work, 128 ;# stack offsets for local variables ;# bottom of stack is cache-aligned for sse use .equiv nb333_ixO, 0.equiv nb333_iyO, 16.equiv nb333_izO, 32.equiv nb333_ixH1, 48.equiv nb333_iyH1, 64.equiv nb333_izH1, 80.equiv nb333_ixH2, 96.equiv nb333_iyH2, 112.equiv nb333_izH2, 128.equiv nb333_ixM, 144.equiv nb333_iyM, 160.equiv nb333_izM, 176.equiv nb333_iqM, 192.equiv nb333_iqH, 208.equiv nb333_dxO, 224.equiv nb333_dyO, 240.equiv nb333_dzO, 256.equiv nb333_dxH1, 272.equiv nb333_dyH1, 288.equiv nb333_dzH1, 304.equiv nb333_dxH2, 320.equiv nb333_dyH2, 336.equiv nb333_dzH2, 352.equiv nb333_dxM, 368.equiv nb333_dyM, 384.equiv nb333_dzM, 400.equiv nb333_qqM, 416.equiv nb333_qqH, 432.equiv nb333_rinvO, 448.equiv nb333_rinvH1, 464.equiv nb333_rinvH2, 480.equiv nb333_rinvM, 496.equiv nb333_rO, 512.equiv nb333_rH1, 528.equiv nb333_rH2, 544.equiv nb333_rM, 560.equiv nb333_tsc, 576.equiv nb333_two, 592.equiv nb333_c6, 608.equiv nb333_c12, 624.equiv nb333_vctot, 640.equiv nb333_Vvdwtot, 656.equiv nb333_fixO, 672.equiv nb333_fiyO, 688.equiv nb333_fizO, 704.equiv nb333_fixH1, 720.equiv nb333_fiyH1, 736.equiv nb333_fizH1, 752.equiv nb333_fixH2, 768.equiv nb333_fiyH2, 784.equiv nb333_fizH2, 800.equiv nb333_fixM, 816.equiv nb333_fiyM, 832.equiv nb333_fizM, 848.equiv nb333_fjx, 864.equiv nb333_fjy, 880.equiv nb333_fjz, 896.equiv nb333_half, 912.equiv nb333_three, 928.equiv nb333_is3, 944.equiv nb333_ii3, 948.equiv nb333_ntia, 952.equiv nb333_innerjjnr, 956.equiv nb333_innerk, 960.equiv nb333_n, 964.equiv nb333_nn1, 968.equiv nb333_nri, 972.equiv nb333_nouter, 976.equiv nb333_ninner, 980.equiv nb333_salign, 984 push ebp mov ebp,esp push eax push ebx push ecx push edx push esi push edi sub esp, 988 ;# local stack space mov eax, esp and eax, 0xf sub esp, eax mov [esp + nb333_salign], eax emms ;# Move args passed by reference to stack mov ecx, [ebp + nb333_p_nri] mov ecx, [ecx] mov [esp + nb333_nri], ecx ;# zero iteration counters mov eax, 0 mov [esp + nb333_nouter], eax mov [esp + nb333_ninner], eax mov eax, [ebp + nb333_p_tabscale] movss xmm5, [eax] shufps xmm5, xmm5, 0 movaps [esp + nb333_tsc], xmm5 ;# create constant floating-point factors on stack mov eax, 0x3f000000 ;# constant 0.5 in IEEE (hex) mov [esp + nb333_half], eax movss xmm1, [esp + nb333_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 + nb333_half], xmm1 movaps [esp + nb333_two], xmm2 movaps [esp + nb333_three], xmm3 ;# assume we have at least one i particle - start directly mov ecx, [ebp + nb333_iinr] ;# ecx = pointer into iinr[] mov ebx, [ecx] ;# ebx =ii mov edx, [ebp + nb333_charge] movss xmm4, [edx + ebx*4 + 4] movss xmm3, [edx + ebx*4 + 12] mov esi, [ebp + nb333_p_facel] movss xmm5, [esi] mulss xmm3, xmm5 mulss xmm4, xmm5 shufps xmm3, xmm3, 0 shufps xmm4, xmm4, 0 movaps [esp + nb333_iqM], xmm3 movaps [esp + nb333_iqH], xmm4 mov edx, [ebp + nb333_type] mov ecx, [edx + ebx*4] shl ecx, 1 mov edi, [ebp + nb333_p_ntype] imul ecx, [edi] ;# ecx = ntia = 2*ntype*type[ii0] mov [esp + nb333_ntia], ecx .nb333_threadloop: mov esi, [ebp + nb333_count] ;# pointer to sync counter mov eax, [esi].nb333_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 .nb333_spinlock ;# if(nn1>nri) nn1=nri mov ecx, [esp + nb333_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 + nb333_n], eax mov [esp + nb333_nn1], ebx sub ebx, eax ;# calc number of outer lists mov esi, eax ;# copy n to esi jg .nb333_outerstart jmp .nb333_end .nb333_outerstart: ;# ebx contains number of outer iterations add ebx, [esp + nb333_nouter] mov [esp + nb333_nouter], ebx.nb333_outer: mov eax, [ebp + nb333_shift] ;# eax = pointer into shift[] mov ebx, [eax + esi*4] ;# ebx=shift[n] lea ebx, [ebx + ebx*2] ;# ebx=3*is mov [esp + nb333_is3],ebx ;# store is3 mov eax, [ebp + nb333_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 + nb333_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 + nb333_pos] ;# eax = base of pos[] mov [esp + nb333_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 + nb333_ixO], xmm3 movaps [esp + nb333_iyO], xmm4 movaps [esp + nb333_izO], xmm5 movaps [esp + nb333_ixH1], xmm6 movaps [esp + nb333_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 + nb333_izH1], xmm6 movaps [esp + nb333_ixH2], xmm0 movaps [esp + nb333_iyH2], xmm1 movaps [esp + nb333_izH2], xmm2 movaps [esp + nb333_ixM], xmm3 movaps [esp + nb333_iyM], xmm4 movaps [esp + nb333_izM], xmm5 ;# clear vctot and i forces xorps xmm4, xmm4 movaps [esp + nb333_vctot], xmm4 movaps [esp + nb333_Vvdwtot], xmm4 movaps [esp + nb333_fixO], xmm4 movaps [esp + nb333_fiyO], xmm4 movaps [esp + nb333_fizO], xmm4 movaps [esp + nb333_fixH1], xmm4 movaps [esp + nb333_fiyH1], xmm4 movaps [esp + nb333_fizH1], xmm4 movaps [esp + nb333_fixH2], xmm4 movaps [esp + nb333_fiyH2], xmm4 movaps [esp + nb333_fizH2], xmm4 movaps [esp + nb333_fixM], xmm4 movaps [esp + nb333_fiyM], xmm4 movaps [esp + nb333_fizM], xmm4 mov eax, [ebp + nb333_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 + nb333_pos] mov edi, [ebp + nb333_faction] mov eax, [ebp + nb333_jjnr] shl ecx, 2 add eax, ecx mov [esp + nb333_innerjjnr], eax ;# pointer to jjnr[nj0] mov ecx, edx sub edx, 4 add ecx, [esp + nb333_ninner] mov [esp + nb333_ninner], ecx add edx, 0 mov [esp + nb333_innerk], edx ;# number of innerloop atoms jge .nb333_unroll_loop jmp .nb333_odd_inner.nb333_unroll_loop: ;# quad-unroll innerloop here mov edx, [esp + nb333_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 + nb333_innerjjnr], 16 ;# advance pointer (unrolled 4) mov esi, [ebp + nb333_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 + nb333_iqM] mulps xmm4, [esp + nb333_iqH] movd mm0, eax ;# use mmx registers as temp storage movd mm1, ebx movd mm2, ecx movd mm3, edx movaps [esp + nb333_qqM], xmm3 movaps [esp + nb333_qqH], xmm4 mov esi, [ebp + nb333_type] mov eax, [esi + eax*4] mov ebx, [esi + ebx*4] mov ecx, [esi + ecx*4] mov edx, [esi + edx*4] mov esi, [ebp + nb333_vdwparam] shl eax, 1 shl ebx, 1 shl ecx, 1 shl edx, 1 mov edi, [esp + nb333_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 + nb333_c6], xmm4 movaps [esp + nb333_c12], xmm6 mov esi, [ebp + nb333_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 + nb333_ixO] movaps xmm5, [esp + nb333_iyO] movaps xmm6, [esp + nb333_izO] ;# calc dr subps xmm4, xmm0 subps xmm5, xmm1 subps xmm6, xmm2 ;# store dr movaps [esp + nb333_dxO], xmm4 movaps [esp + nb333_dyO], xmm5 movaps [esp + nb333_dzO], xmm6 ;# square it mulps xmm4,xmm4 mulps xmm5,xmm5 mulps xmm6,xmm6 addps xmm4, xmm5 addps xmm4, xmm6 movaps xmm7, xmm4 ;# rsqO in xmm7⌨️ 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?